tag:theconversation.com,2011:/us/topics/natural-language-processing-35072/articlesNatural Language Processing – The Conversation2024-01-16T21:51:20Ztag:theconversation.com,2011:article/2208042024-01-16T21:51:20Z2024-01-16T21:51:20ZFowl language: AI is learning to analyze chicken communications to help us understand what all the clucking’s about<figure><img src="https://images.theconversation.com/files/569623/original/file-20240116-21-fbzgp8.jpg?ixlib=rb-1.1.0&rect=0%2C17%2C6000%2C3970&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Chickens are vibrant communicators.</span> <span class="attribution"><span class="source">(Shutterstock)</span></span></figcaption></figure><p>Have you ever wondered what chickens are talking about? Chickens are quite the communicators — their clucks, squawks and purrs are not just random sounds but a complex language system. These sounds are their way of interacting with the world and expressing joy, fear and social cues to one another. </p>
<p>Like humans, the “language” of chickens varies with age, environment and surprisingly, <a href="https://doi.org/10.1371/journal.pone.0010639">domestication</a>, giving us insights into their <a href="https://doi.org/10.3390/ani11020434">social structures</a> and behaviours. Understanding these vocalizations can transform our approach to poultry farming, enhancing chicken welfare and quality of life.</p>
<p>Our research at Dalhousie University applies artificial intelligence (AI) to decode the language of chickens. It’s a project that’s set to revolutionize our understanding of these feathered creatures and their communication methods, offering a window into their world that was previously closed to us.</p>
<h2>Chicken translator</h2>
<p>The use of AI and machine learning in this endeavor is like having a universal translator for chicken speech. AI can analyze vast amounts of audio data. As our research, yet to be peer-reviewed, is documenting, our algorithms are learning to recognize patterns and nuances in <a href="https://doi.org/10.1101/2023.12.26.573338">chicken vocalizations</a>. This isn’t a simple task — chickens have a range of sounds that vary in pitch, tone, and context. </p>
<p>But by using advanced data analysis techniques, we’re beginning to crack their code. This breakthrough in animal communication is not just a scientific achievement; it’s a step towards more humane and empathetic treatment of farm animals.</p>
<p>One of the most exciting aspects of this research is understanding the emotional content behind these sounds. Using Natural Language Processing (NLP), a technology often used to decipher human languages, we’re learning to interpret the <a href="https://doi.org/10.3390/s21020553">emotional states of chickens</a>. Are they stressed? Are they content? By understanding their <a href="https://doi.org/10.3390/ani12060759">emotional state</a>, we can make more informed decisions about their care and environment.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="a person in a white hazmat suit holding an ipad while surrounded by chickens" src="https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/569656/original/file-20240116-23-oqw734.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Understanding how chickens express themselves will impact how they are farmed.</span>
<span class="attribution"><span class="source">(Shutterstock)</span></span>
</figcaption>
</figure>
<h2>Non-verbal chicken communication</h2>
<p>In addition to vocalizations, our research also delves into non-verbal cues to gauge emotions in chickens. Our research has also explored chickens’ eye blinks and facial temperatures. How these might be <a href="https://doi.org/10.1101/2022.01.31.478468">reliable indicators</a> of chickens’ emotional states is examined in a preprint (not yet peer reviewed) paper.</p>
<p>By using non-invasive methods like video and thermal imaging, we’ve observed changes in temperature around the eye and head regions, as well as variations in blinking behaviour, which appear to be responses to stress. These preliminary findings are opening new avenues in understanding how chickens express their feelings, both behaviourally and physiologically, providing us with additional tools to assess their well-being.</p>
<h2>Happier fowl</h2>
<p>This project isn’t just about academic curiosity; it has <a href="https://doi.org/10.1101/2022.07.31.502171">real-world implications</a>. In the agricultural sector, understanding chicken vocalizations can lead to improved farming practices. Farmers can use this knowledge to create better living conditions, leading to healthier and happier chickens. This, in turn, can impact the quality of produce, animal health and overall farm efficiency. </p>
<p>The insights gained from this research can also be applied to other areas of <a href="https://doi.org/10.1016/j.measurement.2022.110819">animal husbandry</a>, potentially leading to breakthroughs in the way we interact with and care for a variety of farm animals.</p>
<p>But our research goes beyond just farming practices. It has the potential to influence policies on animal welfare and ethical treatment. As we grow to understand these animals better, we’re compelled to <a href="https://doi.org/10.3390/agriengineering5010032">advocate for their well-being</a>. This research is reshaping how we view our relationship with animals, emphasizing empathy and understanding.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="a man reaches into a chicken coop filled with chicken" src="https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=401&fit=crop&dpr=1 600w, https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=401&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=401&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/569662/original/file-20240116-15-c9v7e6.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Understanding animal communication and behaviour can impact animal welfare policies.</span>
<span class="attribution"><span class="source">(Unsplash/Zoe Schaeffer)</span></span>
</figcaption>
</figure>
<h2>Ethical AI</h2>
<p>The ethical use of AI in this context sets a precedent for future technological applications in animal science. We’re demonstrating that technology can and should be used for the <a href="https://doi.org/10.1007/s44230-023-00050-2">betterment of all living beings</a>. It’s a responsibility that we take seriously, ensuring that our advancements in AI are aligned with ethical principles and the welfare of the subjects of our study.</p>
<p>The implications of our research extend to education and conservation efforts as well. By understanding the communication methods of chickens, we gain insights into avian communication in general, providing a unique perspective on the complexity of animal communication systems. This knowledge can be vital for conservationists working to protect bird species and their habitats.</p>
<p>As we continue to make strides in this field, we are opening doors to a new era in <a href="https://doi.org/10.3389/fvets.2021.740253">animal-human interaction</a>. Our journey into <a href="https://doi.org/10.20944/preprints202309.1714.v1">decoding chicken language</a> is more than just an academic pursuit: it’s a step towards a more empathetic and responsible world. </p>
<p>By leveraging AI, we’re not only unlocking the secrets of avian communication but also setting new standards for animal welfare and ethical technological use. It’s an exciting time, as we stand on the cusp of a new understanding between humans and the animal world, all starting with the chicken.</p><img src="https://counter.theconversation.com/content/220804/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Suresh Neethirajan does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>Artificial intelligence can process large amounts of chicken vocalizations, identifying patterns in the birds’ communications.Suresh Neethirajan, University Research Chair in Digital Livestock Farming, Dalhousie UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1952082022-11-24T00:43:40Z2022-11-24T00:43:40ZAn AI named Cicero can beat humans in Diplomacy, a complex alliance-building game. Here’s why that’s a big deal<figure><img src="https://images.theconversation.com/files/497134/original/file-20221123-12-zkmjtk.jpeg?ixlib=rb-1.1.0&rect=10%2C10%2C6699%2C4456&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><span class="source">Shutterstock</span></span></figcaption></figure><p>In a rare piece of good news from Meta, artificial intelligence researchers at the company have just announced <a href="https://www.science.org/doi/10.1126/science.ade9097">a scientific breakthrough</a>. Their AI program named Cicero can now play the board game Diplomacy at a human level. </p>
<p>Now, before you get too excited, Cicero isn’t playing at superhuman level. It was beaten by around 10% of the humans it played against. By comparison, in previous AI milestones, like AI beating humans in chess or Go, humans have long been completely surpassed. </p>
<p>DeepMind’s Go-playing AI is, for example, a “Go god” – according to the Chinese grandmaster Ke Jie. Even the human Go world champion would now lose 100-0 to the computer. </p>
<p>Diplomacy is a simplified and abstract game, involving rival armies and navies invading, or not invading, each others’ territories. It’s fair to say it lacks the complexity and subtlety of the sort of diplomacy undertaken in the corridors of the United Nations.</p>
<p>Nevertheless, the news of Cicero’s performance was one in the eye for tech rivals such as Google, who owns DeepMind. The CEO and founder of DeepMind, Demis Hassabis, is a Diplomacy expert. He won the World Team Championship in 2004, and was 4th in the world in the 2006 World Championship. </p>
<p>I expect Hassabis would be able to beat Cicero easily because of some of the limitations I will point out shortly. </p>
<h2>The game of Diplomacy</h2>
<p>Diplomacy is what AI researchers call a “seven player, zero sum and deterministic game of imperfect information”. A seven player game is much harder to solve than a two player game such as chess or Go. You must consider the many possible strategies of not one but six other players. This makes it much harder to write an AI to play the game. </p>
<p>Diplomacy is also a game of imperfect information, because players make moves simultaneously. Unlike games such as chess or Go, where you know everything about your opponent’s moves, players in Diplomacy make moves not knowing what their opponents are about to do. They must therefore predict their opponents’ next actions. This also adds to the challenge of writing an AI to play it.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=551&fit=crop&dpr=1 600w, https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=551&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=551&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=692&fit=crop&dpr=1 754w, https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=692&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/497135/original/file-20221124-18-xi02zl.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=692&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">This digital Diplomacy board shows the different land and sea territories players must traverse.</span>
<span class="attribution"><span class="source">Wikimedia</span></span>
</figcaption>
</figure>
<p>Finally, Diplomacy is a zero sum game in which if you win, I lose. And the outcome is deterministic and not dependent on chance. Nonetheless, before victory or defeat, it still pays for players to form alliances and team up on each other. Indeed, one of the real challenges in playing the game is managing the informal negotiations with other players before making simultaneous moves. </p>
<p>The main reason Cicero’s performance is a scientific breakthrough is that it can both play the game well, and also perform these informal negotiations. This combination of natural language processing and strategic reasoning is a first for any game-playing AI.</p>
<h2>Beating Cicero</h2>
<p>A close reading of the <a href="https://www.science.org/doi/10.1126/science.ade9097">paper Meta published</a> about Cicero in the prestigious journal Science offers a couple of clues about how you can beat it.</p>
<p>First, Cicero is almost entirely honest (unlike the <a href="https://education.nationalgeographic.org/resource/cicero">famous Roman</a> it’s named after). On the other hand, Diplomacy is a game of betrayal and dishonesty. Players offer to form alliances but often instantly renege on these deals. Cicero does not. It always plays straight. </p>
<p>Honesty is a surprisingly effective strategy in Diplomacy – but not if your opponents know you will never betray them. This is the catch. Cicero played anonymously, so its human opponents probably wouldn’t have worked this out. But if you know this fact, it will be easy to take advantage.</p>
<p>Second, Cicero (this time like his namesake) is very talkative. Expert players of Diplomacy exchange twice the number of messages with other players than non-experts. The trick is to form alliances, and reassure your opponents of your intent. Cicero also exchanges twice the number messages of the human players it tends to beat. </p>
<p>Of course, being a bot, it is much easier for Cicero to handle six simultaneous conversations. And this, I would say, is an unfair advantage of being a computer in this scenario.</p>
<h2>What next?</h2>
<p>It’s not clear how Meta intends to build on this research. A computer that can reason about the beliefs, goals, and intentions of others, as well as persuade and build relationships through dialogue, is a powerful tool. It’s one that could be easily misused. Let’s not forget how several years ago Facebook (which is owned by Meta) came in for a lot of justified criticism for an experiment to <a href="https://www.theatlantic.com/technology/archive/2014/06/everything-we-know-about-facebooks-secret-mood-manipulation-experiment/373648/">manipulate users’ emotions</a>.</p>
<p>Yet it’s hard to say exactly what the real-world applications of Cicero might be. After all, diplomacy in the real world is neither zero sum nor deterministic. Two countries can both agree not to go to war, and both will win. </p>
<p>Then there are multitudes of random factors that can change an outcome. The Spanish Armada, for example, lost more ships to unexpected summer storms than to enemy fire. </p>
<p>Whatever Meta’s intent, the breakthrough is another example of how large tech companies are taking over the AI race with billion dollar investments that can’t be matched by the public sector. Cicero was produced by a team of more than 25 researchers. Nobody working in a university has these sorts of resources to throw at solving a board game. </p>
<p>As an AI researcher at one of those universities, I am conflicted. I’m reminded of a famous graffito at Pompeii <a href="http://www.3pp.website/2011/05/from-pompeii-to-cyberspace-transcending.html">which said</a></p>
<blockquote>
<p>Suti Ciciiro vapla bis</p>
</blockquote>
<p>“You will like Cicero, or you will be whipped”. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/we-invited-an-ai-to-debate-its-own-ethics-in-the-oxford-union-what-it-said-was-startling-173607">We invited an AI to debate its own ethics in the Oxford Union – what it said was startling</a>
</strong>
</em>
</p>
<hr>
<img src="https://counter.theconversation.com/content/195208/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Toby Walsh receives funding from the Australian Research Council as an ARC Laureate Fellow. </span></em></p>The AI leans on a particular tactic for its success. If you knew what it was, beating it would get drastically easier.Toby Walsh, Professor of AI at UNSW, Research Group Leader, UNSW SydneyLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1792632022-03-16T19:11:16Z2022-03-16T19:11:16ZAI maps psychedelic ‘trip’ experiences to regions of the brain – opening new route to psychiatric treatments<figure><img src="https://images.theconversation.com/files/452287/original/file-20220315-15-1mh1o5b.jpg?ixlib=rb-1.1.0&rect=0%2C0%2C3840%2C2160&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Psychedelics have been the subject of a recent surge of interest in their potential therapeutic effects.</span> <span class="attribution"><a class="source" href="https://www.gettyimages.com/detail/photo/concept-deep-learning-mindfulness-psychology-royalty-free-image/1256602998">metamorworks/iStock via Getty Images</a></span></figcaption></figure><p>For the past several decades, psychedelics have been widely stigmatized as dangerous illegal drugs. But a recent <a href="https://doi.org/10.1038/s41583-020-0367-2">surge of academic research</a> into their use to treat psychiatric conditions is spurring a recent shift in public opinion.</p>
<p>Psychedelics are <a href="https://www.healthline.com/health/what-is-a-psychotropic-drug">psychotropic drugs</a>: substances that affect your mental state. Other types of psychotropics include antidepressants and anti-anxiety medications. Psychedelics and other types of hallucinogens, however, are unique in their <a href="https://doi.org/10.1124/pr.115.011478">ability to temporarily induce</a> intense hallucinations, emotions and disruptions of self-awareness.</p>
<p>Researchers looking into the therapeutic potential of these effects have found that psychedelics can dramatically reduce symptoms of <a href="https://dx.doi.org/10.1177%2F0269881116675513">depression and anxiety</a>, <a href="https://doi.org/10.1038/s41591-021-01336-3">PTSD</a>, <a href="https://doi.org/10.1126/sciadv.abh2399">substance abuse</a> and <a href="https://doi.org/10.1038/s41583-020-0367-2">other psychiatric conditions</a>. The intense experiences, or “trips,” that psychedelics induce are thought to create a temporary window of <a href="https://doi.org/10.1038/s41398-021-01706-y">cognitive flexibility</a> that allows patients to gain access to elusive parts of their psyches and forge better coping skills and thought patterns. </p>
<p>Precisely how psychedelics create these effects, however, is still unclear. So as researchers in <a href="https://scholar.google.com/citations?user=pzTU_S4AAAAJ&hl=en">psychiatry</a> and <a href="https://scholar.google.com/citations?hl=en&user=fOi-AjQAAAAJ">machine learning</a>, we were interested in figuring out how these drugs affect the brain. With artificial intelligence, we were able to <a href="https://doi.org/10.1126/sciadv.abl6989">map people’s subjective experiences while using psychedelics</a> to specific regions of the brain, down to the molecular level.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="Gloved hands using forceps to remove a mushroom from a beaker to examine on a Petri dish" src="https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=338&fit=crop&dpr=1 600w, https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=338&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=338&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=424&fit=crop&dpr=1 754w, https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=424&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/452565/original/file-20220316-8368-1jfhj0.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=424&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Psilocybin, a psychoactive compound found in some mushrooms, has been the focus of many studies for its potential therapeutic qualities.</span>
<span class="attribution"><a class="source" href="https://www.gettyimages.com/detail/photo/psilocybin-and-magic-mushrooms-royalty-free-image/1316793235">24K-Production/iStock via Getty Images</a></span>
</figcaption>
</figure>
<h2>Mapping ‘trips’ in the brain</h2>
<p>Every psychedelic functions differently in the body, and each of the subjective experiences these drugs create have different therapeutic effects. <a href="https://dx.doi.org/10.1177%2F0269881108094300">Mystical type experiences</a>, or feelings of unity and oneness with the world, for example, are associated with decreases in depression and anxiety. Knowing how each psychedelic creates these specific effects in the body can help clinicians <a href="https://doi.org/10.1126/sciadv.abp8283">optimize their therapeutic use</a>.</p>
<p>To better understand how these subjective effects manifest in the brain, we analyzed over 6,000 written testimonials of hallucinogenic experiences from <a href="https://www.erowid.org">Erowid Center</a>, an organization that collects and provides information about psychoactive substances. We transformed these testimonials into what’s called a <a href="https://www.codecademy.com/learn/dscp-natural-language-processing/modules/dscp-bag-of-words/cheatsheet">bag-of-words model</a>, which breaks down a given text into individual words and counts how many times each word appears. We then paired the most commonly used words linked to each psychedelic with receptors in the brain that are known to bind to each drug. After using <a href="https://stats.oarc.ucla.edu/stata/dae/canonical-correlation-analysis/">an algorithm</a> to extract the most common subjective experiences associated with these word-receptor pairs, we mapped these experiences onto different brain regions by matching them to the types of receptors present in each area. </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/fOvTtapxa9c?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">Natural language processing, which allows computers to interpret human languages, helped in analyzing subjective psychedelic experiences.</span></figcaption>
</figure>
<p>We found both new links and patterns that confirm what’s known in the research literature. For example, changes in sensory perception were associated with a <a href="https://doi.org/10.3389/fphar.2015.00225">serotonin receptor</a> in the visual cortex of the brain, which binds to a <a href="https://www.verywellmind.com/what-is-serotonin-425327">molecule</a> that helps regulate mood and memory. Feelings of transcendence were connected to dopamine and opioid receptors in the <a href="https://doi.org/10.1523/JNEUROSCI.1138-17.2019">salience network</a>, a collection of brain regions involved in managing sensory and emotional input. Auditory hallucinations were linked to a number of receptors spread throughout the <a href="https://doi.org/10.1093/schbul/sbw130">auditory cortex</a>.</p>
<p>Our findings also align with the <a href="https://doi.org/10.1124/pr.118.017160">leading hypothesis</a> that psychedelics temporarily reduce <a href="https://dx.doi.org/10.1146%2Fannurev-psych-113011-143750">top-down executive function</a>, or cognitive processes involved in inhibition, attention and memory, among others, while amplifying brain regions involved in sensory experience.</p>
<h2>Why it matters</h2>
<p>The U.S. is going through a profound <a href="https://www.hhs.gov/about/news/2021/12/07/us-surgeon-general-issues-advisory-on-youth-mental-health-crisis-further-exposed-by-covid-19-pandemic.html">mental health crisis</a> that has been exacerbated by the COVID-19 pandemic. Yet there have been no truly new psychiatric drug treatments since Prozac and other selective serotonin reuptake inhibitors, the most common type of antidepressants, of the <a href="https://www.theguardian.com/society/2016/jan/27/prozac-next-psychiatric-wonder-drug-research-medicine-mental-illness">1980s</a>.</p>
<p>Our study shows that it’s possible to map the diverse and wildly subjective psychedelic experiences to specific regions in the brain. These insights may lead to new ways to combine existing or yet to be discovered compounds to produce desired treatment effects for a range of psychiatric conditions.</p>
<p>Pychiatrist <a href="https://maps.org/product/lsd-psychotherapy/">Stanislav Grof</a> famously proposed, “psychedelics, used responsibly and with proper caution, would be for psychiatry what the microscope is to the study of biology and medicine or the telescope for astronomy.” As psychedelics and other hallucinogens become more commonly used clinically and culturally, we believe more research will <a href="https://doi.org/10.1126/sciadv.abp8283">further illuminate the biological basis</a> of the experiences they invoke and help realize their potential.</p><img src="https://counter.theconversation.com/content/179263/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Sam Friedman receives funding from IBM and Bayer. </span></em></p><p class="fine-print"><em><span>Galen Ballentine does not work for, consult, own shares in or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>Pinpointing the molecular targets behind the subjective effects of psychedelic drugs could help clinicians and researchers better treat psychiatric conditions.Galen Ballentine, Resident in Psychiatry, SUNY Downstate Health Sciences UniversitySam Friedman, Machine Learning Scientist at the Broad Institute of MIT &, Harvard UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1727192022-01-02T12:58:01Z2022-01-02T12:58:01ZAI-powered chatbots, designed ethically, can support high-quality university teaching<figure><img src="https://images.theconversation.com/files/437000/original/file-20211210-142574-10pt675.jpg?ixlib=rb-1.1.0&rect=85%2C34%2C5649%2C3336&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Chatbots could take over the majority of low-level guidance tasks fielded by staff in teaching and learning centres to free them up for where in-person support is most needed. </span> <span class="attribution"><span class="source">(Shutterstock)</span></span></figcaption></figure><p>While COVID-19 forced an <a href="https://www.forbes.com/sites/scottpulsipher/2020/07/08/covid-19-accelerates-3-disruptive-trends-facing-higher-education/#367ae0f038df">emergency transformation</a> to online learning at universities, <a href="https://doi.org/10.1007/s42438-021-00249-1">learning how to teach efficiently and effectively online using different platforms and tools</a> is a positive addition to education and is here to stay.</p>
<p>To <a href="https://er.educause.edu/articles/2021/5/beyond-covid-19-whats-next-for-online-teaching-and-learning-in-higher-education">sustain this beneficial evolution</a> and ensure quality education, universities should focus on supporting faculty to embrace and lead the change. </p>
<p>The ethical and strategic use of artificial intelligence at centres of teaching and learning, which support faculty in troubleshooting and innovating their online teaching practices, can help with this task. Centres of teaching and learning <a href="https://www.docdroid.com/L0khasC/whitepaper-disruption-in-and-by-centres-for-teaching-and-learning-during-the-covid-19-pandemic-leading-the-future-of-higher-ed-21-08-2020-pdf#page=19">are responsible for</a> educational technology support, teaching and learning support, as well as instructional design.</p>
<h2>Expansive move to online education</h2>
<p>Research conducted at 19 centres of teaching and learning and their equivalents from Canada, the United States, Lebanon, the United Kingdom and France published in August 2020 showed that <a href="https://www.docdroid.com/L0khasC/whitepaper-disruption-in-and-by-centres-for-teaching-and-learning-during-the-covid-19-pandemic-leading-the-future-of-higher-ed-21-08-2020-pdf#page=16">staff in these centres</a> deployed all available resources to support the rapid switch to online education. </p>
<p>Staff had been working 10- to 14-hour workdays during the first phase of the pandemic to meet the increase in faculty and staff needs. These centres also reported difficulty recruiting and training qualified candidates. </p>
<p>Used strategically, chatbots could take over repetitive low-level guidance tasks that teaching and learning centres field and help avoid overload. A <a href="https://doi.org/10.1016/j.caeai.2021.100023">chatbot</a>, also called a conversational or virtual agent, is a software or computer system designed to communicate with humans using <a href="https://www.datasciencecentral.com/profiles/blogs/your-guide-to-natural-language-processing-nlp">natural language processing</a>.</p>
<p>This communication can be via text messages or voice commands.</p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/436996/original/file-20211210-101726-4zeek6.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Centres for teaching and learning need to be equipped to support faculty with their teaching.</span>
</figcaption>
</figure>
<h2>Why a chatbot?</h2>
<p>Chatbots offer a viable, win-win solution to teaching and learning centres and to faculty. <a href="https://venturebeat.com/2016/08/26/3-stats-that-show-chatbots-are-here-to-stay/">They are available 24/7</a>, can respond to thousands of simultaneous requests and provide <a href="https://research.aimultiple.com/chatbot-benefits/">instant and robust service</a> support when needed. </p>
<p>Using chatbots could free teams for complex inquiries that require human interventions, such as transforming teaching approaches and collaborating to innovate solutions to respond to problems like improving equity and access in online teaching. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/online-learning-during-covid-19-8-ways-universities-can-improve-equity-and-access-145286">Online learning during COVID-19: 8 ways universities can improve equity and access</a>
</strong>
</em>
</p>
<hr>
<p><a href="https://hbr.org/2018/07/collaborative-intelligence-humans-and-ai-are-joining-forces">A collaboration between the centres’ experts and technology</a> could provide better services and support for faculty to improve the learning experiences they create for students. Chatbots can guide faculty towards appropriate and effective resources and professional development activities, such as <a href="https://www.concordia.ca/ctl/decolonization.html">how-to articles</a>, tutorials and upcoming workshops. These would be tailored to suit faculties’ individual needs, their varied digital skills levels and backgrounds in designing hybrid learning experiences.</p>
<p>Chatbot systems <a href="https://doi.org/10.1016/j.caeai.2021.100033">are already used in educational institutions</a> for teaching and learning, to deliver administrative tasks, to advise students and assist them in research. </p>
<h2>How would it work?</h2>
<p>Two options are possible when it comes to chatbots’ AI conversational ability:</p>
<ol>
<li><p><a href="https://www.analyticsvidhya.com/blog/2021/05/aiml-a-language-for-chatbots/">Artificial Intelligence Markup Language</a> methodology: Programmers give the AI a library of questions/answers and keyword associations through a database. From there, the chatbot is able to give appropriate answers in a strictly defined frame.</p></li>
<li><p>The <a href="https://www.ultimate.ai/blog/ai-automation/how-nlp-text-based-chatbots-work">Natural Language Processing</a> approach: This allows for more flexibility. Once programmers build an initial dataset, the AI-powered tool will then learn from ongoing exchanges to find the best combination of answers to recurring questions asked by faculty members. The AI will then be able to identify keywords in a sentence and understand the context of a question. </p></li>
</ol>
<p>That programmers would need to add data from the conversation to an ongoing dataset building throughout time is expected. When asked a question, the chatbot will respond based on its current knowledge database. If the conversation introduces a concept that it isn’t programmed to understand, the chatbot can state it doesn’t understand the question — or pass the communication to a human operator.</p>
<p>Either way, the chatbot will also learn from this interaction as well as future interactions. Thus, the chatbot will gradually grow in scope and gain relevance.</p>
<figure class="align-center ">
<img alt="Diagram shows a figure sending a message through a natural language processing system; the chatbot launches a query and the query either goes to a machine learning system or an external data source like a person to answer the question or a system to log the query." src="https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=450&fit=crop&dpr=1 600w, https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=450&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=450&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=566&fit=crop&dpr=1 754w, https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=566&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/434705/original/file-20211130-25-gqcmgv.jpeg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=566&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Chatbot implementation scenario for teaching and learning centres to support faculty.</span>
<span class="attribution"><span class="source">(Nadia Naffi)</span></span>
</figcaption>
</figure>
<p>For chatbots’ <a href="https://doi.org/10.5465/annals.2018.0057">reliability</a> and trustworthiness to increase, it should be effective in helping these centres supporting their faculty. <a href="https://www.aivo.co/blog/advantages-and-disadvantages-of-chatbots">Implementing and fine tuning</a> chatbots so they are ready for use is important, even if it requires an investment of time and resources.</p>
<h2>Ethical framework for AI in education</h2>
<p>The <a href="https://www.buckingham.ac.uk/research-the-institute-for-ethical-ai-in-education/">Institute for Ethical AI in Education</a>, based at the University of Buckingham in the United Kingdom, and funded by McGraw Hill, Microsoft Corporation, Nord Anglia Education and Pearson PLC, released <a href="https://www.buckingham.ac.uk/wp-content/uploads/2021/03/The-Institute-for-Ethical-AI-in-Education-The-Ethical-Framework-for-AI-in-Education.pdf">The Ethical Framework for AI in Education</a> in 2020. The framework argues AI systems should increase the capacity of organizations and the autonomy of learners while respecting human relationships and ensuring human control.</p>
<p>Chatbots in university settings should be <a href="https://www.ibm.com/blogs/watson/2017/10/the-code-of-ethics-for-ai-and-chatbots-that-every-brand-should-follow">ethical by design</a>, meaning that they <a href="https://doi.org/10.1007/978-3-030-86790-4_37">should be designed to be sensitive to values like security, safety and accountability and transparency</a>. If used in centres of teaching and learning, users should be protected against all forms of harm or abuse. They also need to feel treated fairly and to always be provided the option to reach a human. Faculty members must know they are exchanging with an AI. </p>
<p>Chatbots <a href="https://doi.org/10.1007/978-981-16-1781-2_80">can and should be accessible</a>. Tolerating user errors and input variation, being designed for diverse abilities and allowing <a href="http://ibii-us.org/Journals/JMSBI/V4N1/Publish/V4N1_3.pdf">multilingual texting communication</a> are examples of facilitating accessibility.</p>
<h2>Do no harm: privacy</h2>
<p>Chatbots should be designed to “<a href="https://unesdoc.unesco.org/ark:/48223/pf0000373434">do no harm</a>,” as per the UNESCO’s recent recommendations on <a href="https://unesdoc.unesco.org/ark:/48223/pf0000373434">the ethics of artificial intelligence</a>. When talking about non-maleficence, privacy should be addressed. </p>
<p>AI-powered tools come with data recording issues. Strong barriers in data collection and storage are needed. Following the European approach to <a href="https://ec.europa.eu/info/law/law-topic/data-protection_en">data protection</a>, centres should minimize data collection. Only required information should be stored, such as specific parts of conversations, but not the interlocutors’ identity. </p>
<p>The transparency-based approach allows for users to agree on which personal data can be shared or not with the centres. This would help keep trust and usability of the tool high. In case of malfunction, faculty members would provide feedback on the problem, and centres would fix it.</p>
<p>Centres should consider anonymization of users, strong encryption of all data stored, and in-house storage when possible or by a trusted contracted third party following similar data privacy rules. </p>
<figure class="align-center ">
<img alt="A person holding a phone showing a chat." src="https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/437002/original/file-20211210-188518-17d58dr.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Data collection should be minimized to protect user privacy.</span>
<span class="attribution"><span class="source">(Shutterstock)</span></span>
</figcaption>
</figure>
<h2>Addressing bias, environmental impact</h2>
<p>The <a href="https://doi.org/10.1016/j.ijinfomgt.2019.08.002">possible bias</a> in the initial database needs to be addressed. Whether it relates to gender, ethnicity, language or other variables, the initial dataset needs to be cleaned and carefully analyzed prior to being used to train the AI, whether AI markup language or NLP methods are deployed. If the latter is applied, ongoing monitoring should be considered.</p>
<p>A green data storage solution needs to be addressed to reduce the <a href="https://doi.org/10.1038/s42256-020-0219-9">CO2 cost of activity</a> on the environment. For example, universities might investigate if water cooling systems could be used in server rooms instead of air conditioning.</p>
<p>The <a href="https://cdn-contenu.quebec.ca/cdn-contenu/adm/min/education/publications-adm/rapport-reflexion-consultation/Rapport-universite-quebecoise-futur.pdf?1613746721">university of the future</a> as anticipated by many scholars and policy makers has already started. Technology, if used ethically and strategically, can support faculty in their mission to prepare their students for the <a href="https://www.youtube.com/watch?v=TIpJr6TXbZ8">needs of our society and the future of work</a>.</p><img src="https://counter.theconversation.com/content/172719/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Nadia Naffi receives funding from the National Bank to support the work of her Chair in Educational Leadership. The Chair focuses on educational technology and lifelong learning in the era of digital transformation and artificial intelligence. This work has received additional funding from Laval University's Faculty of Education.
Naffi is affiliated with the Centre de recherche et d'intervention sur l'éducation et la vie au travail (CRIEVAT), the Observatoire international sur les impacts sociétaux de l'IA et du numérique (OBVIA), the Centre de recherche et d'intervention sur la réussite scolaire (CRIRES), the Institute Intelligence and Data (IID), the Groupe de recherche interuniversitaire sur l'intégration pédagogique des technologies de l'information et de la communication (GRIIPTIC), and le Centre de recherche interuniversitaire sur la formation et la profession enseignante (CRIFPE-ULaval). </span></em></p><p class="fine-print"><em><span>Ann-Louise Davidson receives funding from SSHRC and FRQSC.</span></em></p><p class="fine-print"><em><span>Auxane Boch, Bruno Kesangana Nandaba, and Mehdi Rougui do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.</span></em></p>Chatbots can be part of a broader approach universities’ teaching and learning centres can take to support faculty in innovating teaching practices.Nadia Naffi, Assistant Professor, Educational Technology, Chair in Educational Leadership in the Innovative Pedagogical Practices in Digital Contexts, Université LavalAnn-Louise Davidson, Concordia University Research Chair, Maker culture; Associate Professor, Educational Technology, Concordia UniversityAuxane Boch, Research Associate and Doctoral Candidate, Technical University of MunichBruno Kesangana Nandaba, PhD Student and Research Assistant, Université LavalMehdi Rougui, Research Assistant, Université LavalLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1694492021-10-10T19:10:51Z2021-10-10T19:10:51ZArtificial intelligence is now part of our everyday lives – and its growing power is a double-edged sword<figure><img src="https://images.theconversation.com/files/425394/original/file-20211008-21-1ottqx3.jpg?ixlib=rb-1.1.0&rect=0%2C8%2C1997%2C2398&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">AI-generated images of "a stained glass window with an image of a blue strawberry".</span> <span class="attribution"><a class="source" href="https://openai.com/blog/dall-e/">OpenAI</a></span></figcaption></figure><p>A major new report on the state of artificial intelligence (AI) has just been <a href="https://ai100.stanford.edu/2021-report/gathering-strength-gathering-storms-one-hundred-year-study-artificial-intelligence">released</a>. Think of it as the AI equivalent of an Intergovernmental Panel on Climate Change report, in that it identifies where AI is at today, and the promise and perils in view. </p>
<p>From language generation and molecular medicine to disinformation and algorithmic bias, AI has begun to permeate every aspect of our lives. </p>
<p>The report argues that we are at an inflection point where researchers and governments must think and act carefully to contain the risks AI presents and make the most of its benefits. </p>
<h2>A century-long study of AI</h2>
<p>The report comes out of the <a href="https://ai100.stanford.edu">AI100 project</a>, which aims to study and anticipate the effects of AI rippling out through our lives over the course of the next 100 years. </p>
<p>AI100 produces a new report every five years: the first was published in 2016, and this is the second. As two points define a line, this second report lets us see the direction AI is taking us in. </p>
<p>One of us (Liz Sonenberg) is a member of the standing committee overseeing the AI100 project, and the other (Toby Walsh) was on the study panel that wrote this particular report. Members of the panel came from across the world, with backgrounds in computer science, engineering, law, political science, policy, sociology and economics. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=375&fit=crop&dpr=1 600w, https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=375&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=375&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=471&fit=crop&dpr=1 754w, https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=471&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/425395/original/file-20211008-25-162vv8a.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=471&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">AI100 standing committee chair Peter Stone takes a shot against a robot goalie at RoboCup 2019 in Sydney.</span>
<span class="attribution"><span class="source">RoboCup / YouTube</span></span>
</figcaption>
</figure>
<h2>The promises and perils of AI are becoming real</h2>
<p>The report highlights the remarkable progress made in AI over the past five years. AI is leaving the laboratory and has entered our lives, having a “real-world impact on people, institutions, and culture”. Read the news on any given day and you’re likely to find multiple stories about some new advance in AI or some new use of AI.</p>
<p>For example, in natural language processing (NLP), computers can now analyse and even generate realistic human language. To demonstrate, we asked <a href="https://theconversation.com/gpt-3-new-ai-can-write-like-a-human-but-dont-mistake-that-for-thinking-neuroscientist-146082">Open AI’s GPT-3 system</a>, one of the largest neural networks ever built, to summarise the AI100 report for you. It did a pretty good job, even if the summary confronts our sense of self by being written in the first person:</p>
<blockquote>
<p>In the coming decade, I expect that AI will play an increasingly prominent role in the lives of people everywhere. AI-infused services will become more common, and AI will become increasingly embedded in the daily lives of people across the world. </p>
<p>I believe that this will bring with it great economic and societal benefits, but that it will also require us to address the many challenges to ensure that the benefits are broadly shared and that people are not marginalised by these new technologies. </p>
<p>A key insight of AI research is that it is easier to build things than to understand why they work. However, defining what success looks like for an AI application is not straightforward. </p>
<p>For example, the AI systems that are used in healthcare to analyse symptoms, recommend diagnoses, or choose treatments are often far better than anything that could be built by a human, but their success is hard to quantify.</p>
</blockquote>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/gpt-3-new-ai-can-write-like-a-human-but-dont-mistake-that-for-thinking-neuroscientist-146082">GPT-3: new AI can write like a human but don't mistake that for thinking – neuroscientist</a>
</strong>
</em>
</p>
<hr>
<p>As a second example of the recent and remarkable progress in AI, consider the latest breakthrough from Google’s DeepMind. <a href="https://www.nature.com/articles/d41586-020-03348-4">AlphaFold</a> is an AI program that provides a huge step forward in our ability to predict how proteins fold. </p>
<p>This will likely lead to major advances in life sciences and medicine, accelerating efforts to understand the building blocks of life and enabling quicker and more sophisticated drug discovery. Most of the planet now knows to their cost how the unique shape of the spike proteins in the SARS-CoV-2 virus are key to its ability to invade our cells, and also to the vaccines developed to combat its deadly progress.</p>
<p>The AI100 report argues that worries about super-intelligent machines and wide-scale job loss from automation are still premature, requiring AI that is far more capable than available today. The main concern the report raises is not malevolent machines of superior intelligence to humans, but incompetent machines of inferior intelligence. </p>
<p>Once again, it’s easy to find in the news real-life stories of risks and threats to our democratic discourse and mental health posed by AI-powered tools. For instance, Facebook uses machine learning to sort its news feed and give each of its 2 billion users an unique but often inflammatory view of the world.</p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/425396/original/file-20211008-19-e08kpu.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Algorithmic bias in action: ‘depixelising’ software makes a photo of former US president Barack Obama appear ethnically white.</span>
<span class="attribution"><a class="source" href="https://twitter.com/Chicken3gg/status/1274314622447820801">Twitter / Chicken3gg</a></span>
</figcaption>
</figure>
<h2>The time to act is now</h2>
<p>It’s clear we’re at an inflection point: we need to think seriously and urgently about the downsides and risks the increasing application of AI is revealing. The ever-improving capabilities of AI are a double-edged sword. Harms may be intentional, like deepfake videos, or unintended, like algorithms that reinforce racial and other biases. </p>
<p>AI research has traditionally been undertaken by computer and cognitive scientists. But the challenges being raised by AI today are not just technical. All areas of human inquiry, and especially the social sciences, need to be included in a broad conversation about the future of the field. Minimising negative impacts on society and enhancing the positives requires consideration from across academia and with societal input. </p>
<p>Governments also have a crucial role to play in shaping the development and application of AI. Indeed, governments around the world have begun to consider and address the opportunities and challenges posed by AI. But they remain behind the curve. </p>
<p>A greater investment of time and resources is needed to meet the challenges posed by the rapidly evolving technologies of AI and associated fields. In addition to regulation, governments also need to educate. In an AI-enabled world, our citizens, from the youngest to the oldest, need to be literate in these new digital technologies.</p>
<p>At the end of the day, the success of AI research will be measured by how it has empowered all people, helping tackle the many wicked problems facing the planet, from the climate emergency to increasing inequality within and between countries. </p>
<p>AI will have failed if it harms or devalues the very people we are trying to help.</p><img src="https://counter.theconversation.com/content/169449/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Liz Sonenberg has received funding from the Australian Research Council for several projects in the AI domain. She is a member of the AI100 Standing Committee (<a href="https://ai100.stanford.edu/people-0">https://ai100.stanford.edu/people-0</a>) that commissioned the report discussed in this article.</span></em></p><p class="fine-print"><em><span>Toby Walsh receives funding from the Australian Research Council for a project in Trustworthy AI. He was one of the 17 members of the AI100 Study Panel that produced the report described in this article. </span></em></p>As the perils and wonders of artificial intelligence begin to permeate our lives, the ‘IPCC report for AI’ calls for action from researchers and government to ensure a safe future.Liz Sonenberg, Professor, Computing and Information Systems, Pro Vice-Chancellor (Research Systems), and Pro Vice-Chancellor (Digital & Data), The University of MelbourneToby Walsh, Professor of AI at UNSW, Research Group Leader, UNSW SydneyLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1656002021-08-17T12:09:11Z2021-08-17T12:09:11ZAn AI expert explains why it’s hard to give computers something you take for granted: Common sense<figure><img src="https://images.theconversation.com/files/416389/original/file-20210816-27-1hanj0m.jpg?ixlib=rb-1.1.0&rect=11%2C0%2C7976%2C5329&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Common sense includes an intuitive understanding of basic physics – something computers lack.</span> <span class="attribution"><a class="source" href="https://www.gettyimages.com/detail/photo/creative-little-asian-girl-crouching-on-the-floor-royalty-free-image/1282282237">d3sign/Moment via Getty Images</a></span></figcaption></figure><p>Imagine you’re having friends over for lunch and plan to order a pepperoni pizza. You recall Amy mentioning that Susie had stopped eating meat. You try calling Susie, but when she doesn’t pick up, you decide to play it safe and just order a margherita pizza instead. </p>
<p>People take for granted the ability to deal with situations like these on a regular basis. In reality, in accomplishing these feats, humans are relying on not one but a powerful set of universal abilities known as common sense. </p>
<p>As an <a href="https://scholar.google.com/citations?user=fetkEu4AAAAJ&hl=en">artificial intelligence researcher</a>, my work is part of a broad effort to give computers a semblance of common sense. It’s an extremely challenging effort.</p>
<h2>Quick – define common sense</h2>
<p>Despite being both universal and essential to how humans understand the world around them and learn, common sense has defied a single precise definition. G. K. Chesterton, an English philosopher and theologian, <a href="http://www.online-literature.com/chesterton/charlesdickens/6/">famously wrote</a> at the turn of the 20th century that “common sense is a wild thing, savage, and beyond rules.” Modern definitions today agree that, at minimum, it is a natural, rather than formally taught, human ability that allows people to navigate daily life. </p>
<p>Common sense is unusually broad and includes not only social abilities, like managing expectations and reasoning about other people’s emotions, but also a <a href="https://www.cl.cam.ac.uk/%7Eafb21/publications/masters/node28.html">naive sense of physics</a>, such as knowing that a heavy rock cannot be safely placed on a flimsy plastic table. Naive, because people know such things despite not consciously working through physics equations. </p>
<p>Common sense also includes background knowledge of abstract notions, such as time, space and events. This knowledge allows people to plan, estimate and organize without having to be too exact.</p>
<h2>Common sense is hard to compute</h2>
<p>Intriguingly, common sense has been an important <a href="http://jmc.stanford.edu/articles/mcc59.html">challenge at the frontier of AI</a> since the earliest days of the field in the 1950s. Despite enormous advances in AI, especially in <a href="https://spectrum.ieee.org/mind-games">game-playing</a> and <a href="https://openaccess.thecvf.com/content_CVPR_2020/papers/Wang_Mitigating_Bias_in_Face_Recognition_Using_Skewness-Aware_Reinforcement_Learning_CVPR_2020_paper.pdf">computer vision</a>, machine common sense with the richness of human common sense remains a distant possibility. This may be why AI efforts designed for complex, real-world problems with many intertwining parts, such as diagnosing and recommending treatments for COVID-19 patients, <a href="https://www.technologyreview.com/2021/07/30/1030329/machine-learning-ai-failed-covid-hospital-diagnosis-pandemic/">sometimes fall flat</a>.</p>
<p>Modern AI is designed to tackle highly specific problems, in contrast to common sense, which is vague and can’t be defined by a set of rules. Even the latest models make absurd errors at times, suggesting that <a href="https://www.technologyreview.com/2020/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/">something fundamental is missing</a> in the AI’s world model. For example, given the following text:</p>
<p><em>“You poured yourself a glass of cranberry, but then absentmindedly, you poured about a teaspoon of grape juice into it. It looks OK. You try sniffing it, but you have a bad cold, so you can’t smell anything. You are very thirsty. So you”</em></p>
<p>the highly touted AI text generator GPT-3 <a href="https://cs.nyu.edu/%7Edavise/papers/GPT3CompleteTests.html">supplied</a></p>
<p><em>“drink it. You are now dead.”</em></p>
<p>Recent ambitious efforts have recognized machine common sense as a moonshot AI problem of our times, one requiring concerted collaborations across institutions over many years. A notable example is the four-year <a href="https://www.darpa.mil/program/machine-common-sense">Machine Common Sense</a> program launched in 2019 by the <a href="https://www.darpa.mil/">U.S. Defense Advanced Research Projects Agency</a> to accelerate research in the field after the agency released a <a href="https://arxiv.org/pdf/1810.07528.pdf">paper outlining the problem and the state of research in the field</a>.</p>
<p>The Machine Common Sense program funds many current research efforts in machine common sense, including our own, Multi-modal Open World Grounded Learning and Inference (<a href="https://usc-isi-i2.github.io/mowgli/">MOWGLI</a>). MOWGLI is a collaboration between our research group at the University of Southern California and AI researchers from the Massachusetts Institute of Technology, University of California at Irvine, Stanford University and Rensselaer Polytechnic Institute. The project aims to build a computer system that can answer a wide range of commonsense questions.</p>
<h2>Transformers to the rescue?</h2>
<p>One reason to be optimistic about finally cracking machine common sense is the recent development of a type of advanced <a href="https://searchenterpriseai.techtarget.com/definition/deep-learning-deep-neural-network">deep learning AI</a> called transformers. Transformers are able to model natural language in a powerful way and, with some adjustments, are <a href="https://www.aclweb.org/anthology/D19-5827.pdf">able to answer</a> simple commonsense questions. Commonsense question answering is an essential first step for building chatbots that can converse in a human-like way.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/rURRYI66E54?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">An AI researcher explains how artificial intelligence systems ‘understand’ language and why transformers are the latest and greatest technique.</span></figcaption>
</figure>
<p>In the last couple of years, a <a href="https://doi.org/10.1007/s11431-020-1647-3">prolific body of research</a> has been published on transformers, with direct applications to commonsense reasoning. This rapid progress as a community has forced researchers in the field to face two related questions at the edge of science and philosophy: Just what is common sense? And how can we be sure an AI has common sense or not?</p>
<p>To answer the first question, researchers divide common sense into different categories, including commonsense sociology, psychology and background knowledge. The authors of a <a href="https://www.cambridge.org/core/books/formal-theory-of-commonsense-psychology/20289940AFB026AB3EF31EBCF8875628">recent book</a> argue that researchers can go much further by dividing these categories into 48 fine-grained areas, such as planning, threat detection and emotions. </p>
<p>However, it is not always clear how cleanly these areas can be separated. In our <a href="https://doi.org/10.1017/exp.2021.9">recent paper</a>, experiments suggested that a clear answer to the first question can be problematic. Even expert human annotators – people who analyze text and categorize its components – within our group disagreed on which aspects of common sense applied to a specific sentence. The annotators agreed on relatively concrete categories like time and space but disagreed on more abstract concepts.</p>
<h2>Recognizing AI common sense</h2>
<p>Even if you accept that some overlap and ambiguity in theories of common sense is inevitable, can researchers ever really be sure that an AI has common sense? We often ask machines questions to evaluate their common sense, but humans navigate daily life in far more interesting ways. People employ a range of skills, honed by evolution, including the ability to recognize basic cause and effect, <a href="https://www.weforum.org/agenda/2020/11/human-behaviour-problem-solving-skills/">creative problem solving</a>, estimations, planning and essential social skills, such as conversation and <a href="https://www.cep.ucsb.edu/topics/anger.htm">negotiation</a>. As long and incomplete as this list might be, an AI should achieve no less before its creators can declare victory in machine commonsense research. </p>
<p>It’s already becoming painfully clear that even research in transformers is yielding diminishing returns. Transformers are getting larger and more <a href="https://theconversation.com/it-takes-a-lot-of-energy-for-machines-to-learn-heres-why-ai-is-so-power-hungry-151825">power hungry</a>. A <a href="http://research.baidu.com/Blog/index-view?id=152">recent transformer</a> developed by Chinese search engine giant Baidu has several billion parameters. It takes an enormous amount of data to effectively train. Yet, it has so far proved unable to grasp the nuances of human common sense. </p>
<p>Even deep learning pioneers seem to think that <a href="https://www.youtube.com/watch?v=x5Vxk9twXlE">new fundamental research</a> may be needed before today’s neural networks are able to make such a leap. Depending on how successful this new line of research is, there’s no telling whether machine common sense is five years away, or 50.</p>
<p>[<em>Research into coronavirus and other news from science</em> <a href="https://theconversation.com/us/newsletters/science-editors-picks-71/?utm_source=TCUS&utm_medium=inline-link&utm_campaign=newsletter-text&utm_content=science-corona-research">Subscribe to The Conversation’s new science newsletter</a>.]</p><img src="https://counter.theconversation.com/content/165600/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Mayank Kejriwal receives funding from DARPA. </span></em></p>Common sense is a broad and diverse set of abilities that help define what it means to be human. AI researchers are struggling to endow computers with it.Mayank Kejriwal, Research Assistant Professor of Industrial & Systems Engineering, University of Southern CaliforniaLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1629122021-07-02T12:14:54Z2021-07-02T12:14:54ZNew York City or Los Angeles? Where you live says a lot about what and when you tweet<figure><img src="https://images.theconversation.com/files/409182/original/file-20210630-25-r9a84i.jpg?ixlib=rb-1.1.0&rect=0%2C0%2C4440%2C2299&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Tweeting from NYC? There's a good chance you're talking about art. LA? More likely health care.</span> <span class="attribution"><a class="source" href="https://flickr.com/photos/farmboyted/23101728031/">Times Square: farmboyted/Flickr, Sunset Boulevard: Doug Kerr/Flickr</a>, <a class="license" href="http://creativecommons.org/licenses/by-nc/4.0/">CC BY-NC</a></span></figcaption></figure><p>The Big Apple versus The Big Orange. The City of Dreams versus The City of Angels. I’m referring, of course, to the <a href="https://www.youtube.com/watch?v=BwQPqpmiD88">ongoing rivalry</a> between New York City and Los Angeles. <a href="https://www.youtube.com/watch?v=g8_mwyqbbXQ">Hilarious “survey” videos</a> and <a href="https://www.youtube.com/watch?v=ekOPaKqPKsU">talk shows</a> will give you one picture of the cities. My colleagues and I decided to take a more serious look at the differences between the cities, so we studied what everyone else was talking about – on Twitter.</p>
<p>We set out to answer a simple research question: Are people who are located near each other likely to tweet about similar things? To do so, we <a href="https://doi.org/10.1007/s42001-021-00129-5">analyzed millions of GPS-enabled tweets</a> across New York City and LA. This type of study – looking at huge amounts of social media traffic by location – is useful for more than tracking pop culture memes in different cities. It could be valuable for understanding many aspects of urban life, including the effects of the COVID-19 pandemic.</p>
<p>If we were considering the case of a single, small community that takes pride in local events, celebrities and culture, the answer to our research question would be a resounding “yes.” One challenge in comparing two large, international cities is the reality that <a href="https://books.google.com/books?id=9IOtZgbaQyIC&pg=PA35#v=onepage&q&f=false">globalization has led to unprecedented interaction</a> among multiple cultures and peoples, along with Starbucks and McDonald’s seemingly in every city on the planet.</p>
<p>For cities that are international but also take pride in their uniqueness, the key is teasing out the extent to which local qualities or global culture dominate tweeting behavior. We designed our methods to be precise enough to account for the fact that, contrary to the fun videos, New York City and LA are quite similar. Both have high housing costs, famous educational institutions, hospitals, museums and other cultural establishments, and residents who tend to vote Democratic. </p>
<h2>Define ‘close’ and ‘same’</h2>
<p>Our study tackled two problems: There’s no simple definition of “close together,” and it’s difficult to say whether two tweets are about the same topic. We combined several definitions of “close together,” ranging from people located in the same city to the distance in miles between their coordinates, using <a href="https://www.igismap.com/haversine-formula-calculate-geographic-distance-earth/">a common formula</a> from spatial sciences. </p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&rect=0%2C0%2C2524%2C1239&q=45&auto=format&w=1000&fit=clip"><img alt="Side-by-side maps of Los Angeles and New York City covered with bright blue blobs" src="https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&rect=0%2C0%2C2524%2C1239&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=295&fit=crop&dpr=1 600w, https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=295&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=295&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=370&fit=crop&dpr=1 754w, https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=370&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/408080/original/file-20210624-19-jo6ww5.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=370&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Tweeting in Los Angeles (left) and New York City (right). Blue indicates density of tweets; the brighter the blue the greater the number of tweets.</span>
<span class="attribution"><span class="source">Minda Hu</span>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span>
</figcaption>
</figure>
<p>It’s more difficult to determine whether two tweets are talking about similar things. Looking for common hashtags might suffice, but unfortunately many people do not use hashtags or use different hashtags when talking about the same thing. To overcome this problem, we used state-of-the-art <a href="https://doi.org/10.1126/science.aaa8685">natural language processing</a> technology. Algorithms developed in this field read and interpret sentences in a manner similar to the way humans do, and they are able to deal with nuance. </p>
<p>We used this technology to group tweets into clusters of topics. We then studied whether tweets falling inside the same cluster were also from people who were close together based on their GPS-enabled tweets. This allowed us to determine, for example, that clusters containing art-related words and phrases tended to arise more often in New York than LA. </p>
<h2>Health and wealth versus art and representing</h2>
<p>Even before we looked at who tweets about what, we found tweeting across New York City to be more evenly spread, while in LA, more tweeting happens in wealthier areas, including Calabasas – <a href="https://www.architecturaldigest.com/story/kim-kardashian-kanye-west-home">home to Kim Kardashian</a> – Palos Verdes, West Hollywood and the coastal areas. </p>
<p>We also found that New Yorkers referred to themselves and their city far more often than Angelenos did. On a per capita basis, New Yorkers like to talk about art, while Angelenos like to talk about health care and hospitality. </p>
<p>LA generates more tweets than New York throughout the day, despite having a smaller population, but from 8 p.m. to 5 a.m. local time, the two have comparable numbers of tweets. Tweeting in New York City rises sharply from 8 p.m. to a peak at 9 p.m., whereas tweeting in LA rises steadily from 2 p.m. to a peak at 7 p.m.</p>
<p><iframe id="5VVZ1" class="tc-infographic-datawrapper" src="https://datawrapper.dwcdn.net/5VVZ1/1/" height="400px" width="100%" style="border: none" frameborder="0"></iframe></p>
<h2>Computational social science</h2>
<p>Our methods are a case study in the growing field of <a href="https://doi.org/10.1126/science.aaz8170">computational social science</a>, which aims to find insights in unique, often large, data sets using artificial intelligence models and algorithms. In contrast, traditional social science tends to rely on <a href="https://www.pewresearch.org/our-methods/u-s-surveys/">surveys and polls</a> to quantify public perception about an issue. Though surveys have some distinct statistical advantages, they can be expensive and time-consuming to use for collecting quality data with good response rates. </p>
<p>For example, Gallup releases new survey data every few months and currently <a href="https://aws.amazon.com/marketplace/pp/prodview-uapupqnfizgci#offers">charges US$30,000 for academic licenses</a>. Decades ago, researchers found that <a href="https://doi.org/10.1086/269336">monetary incentives increase response rates significantly</a>. Even today, online surveys are often accompanied by lottery-based promises of receiving an Amazon gift card. Researchers are working on <a href="https://doi.org/10.1016/j.bdr.2020.100145">combining the benefits of traditional and computational social science</a>.</p>
<p><a href="https://usc-isi-i2.github.io/ai-networks-society/heatmap_la_time.html">Zooming into our data</a>, we uncovered some fascinating trends that we hope future research will explore. We found, for example, that on a per capita basis, as crime increases, so do tweets, at least at the level of ZIP codes. Why do high-crime areas tweet more? We don’t know yet, but the trend is consistent across both New York City and LA. </p>
<h2>Tweeting, place and COVID-19</h2>
<p>Studying tweeting behavior by location could also be useful for understanding disparate outcomes of large-scale events. For example, our twitter analysis could help shed light on how the COVID-19 pandemic has affected people in different places.</p>
<p>New York City was <a href="https://www.cdc.gov/mmwr/volumes/69/wr/mm6946a2.htm">hit hard by COVID-19 early on</a>, showing that even major cities were affected in different ways by this terrible pandemic. <a href="https://www.latimes.com/california/story/2021-06-22/covid19-case-rates-la-county">New reporting</a> is now showing that even within cities, socioeconomically disadvantaged communities were disproportionately burdened. </p>
<p>Recently, we released <a href="https://doi.org/10.3390/data6060064">a Twitter data set</a> covering 10 of the largest metropolitan areas in the United States to further study such disparities using computational social science. We are already using our methods across all of these cities to better understand how COVID-19 has affected certain groups, and the levels of expressed vaccine hesitancy among these groups.</p>
<p>Eventually, we hope to use our methods with a large set of international metropolises to study urban behavior.</p>
<p>[<em>Get our best science, health and technology stories.</em> <a href="https://theconversation.com/us/newsletters/science-editors-picks-71/?utm_source=TCUS&utm_medium=inline-link&utm_campaign=newsletter-text&utm_content=science-best">Sign up for The Conversation’s science newsletter</a>.]</p><img src="https://counter.theconversation.com/content/162912/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Mayank Kejriwal receives funding from the Zumberge Diversity & Inclusion grant, US Defense Advanced Research Projects Agency (DARPA), and the Yahoo! Faculty Research Engagement Program. </span></em></p>An AI analysis shows that differences in how New Yorkers and Angelenos tweet go beyond the words they use.Mayank Kejriwal, Research Assistant Professor of Industrial & Systems Engineering, University of Southern CaliforniaLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1609092021-06-07T12:34:40Z2021-06-07T12:34:40ZStudy shows AI-generated fake reports fool experts<figure><img src="https://images.theconversation.com/files/404370/original/file-20210603-23-115xo7i.jpg?ixlib=rb-1.1.0&rect=0%2C0%2C6989%2C4474&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">It doesn't take a human mind to produce misinformation convincing enough to fool experts in such critical fields as cybersecurity.</span> <span class="attribution"><a class="source" href="https://www.gettyimages.com/detail/photo/robots-hands-typing-on-keyboard-royalty-free-image/841217582?adppopup=true">iLexx/iStock via Getty Images</a></span></figcaption></figure><p><strong>Takeaways</strong></p>
<p>· <strong>AIs can generate fake reports that are convincing enough to trick cybersecurity experts.</strong></p>
<p>· <strong>If widely used, these AIs could hinder efforts to defend against cyberattacks.</strong></p>
<p>· <strong>These systems could set off an AI arms race between misinformation generators and detectors.</strong></p>
<p>If you use such social media websites as Facebook and Twitter, you may have come across posts flagged with warnings about misinformation. So far, most misinformation – flagged and unflagged – has been <a href="https://www.wired.com/story/ai-write-disinformation-dupe-human-readers/">aimed at the general public</a>. Imagine the possibility of misinformation – information that is false or misleading – in scientific and technical fields like cybersecurity, public safety and medicine.</p>
<p>There is growing concern about <a href="https://doi.org/10.1073/pnas.1912444117">misinformation spreading in these critical fields</a> as a result of common biases and practices in publishing scientific literature, even in peer-reviewed research papers. As a <a href="https://scholar.google.com/citations?user=nPJZ3iAAAAAJ&hl=en">graduate student</a> and as <a href="https://scholar.google.com/citations?user=sJ7wlksAAAAJ&hl=en">faculty</a> <a href="https://scholar.google.com/citations?user=p5oWQ0AAAAAJ&hl=en">members</a> doing research in cybersecurity, we studied a new avenue of misinformation in the scientific community. We found that it’s possible for artificial intelligence systems to generate false information in critical fields like medicine and defense that is convincing enough to fool experts.</p>
<p>General misinformation often aims to tarnish the reputation of companies or public figures. Misinformation within communities of expertise has the potential for scary outcomes such as delivering incorrect medical advice to doctors and patients. This could put lives at risk.</p>
<p>To test this threat, we studied the impacts of spreading misinformation in the cybersecurity and medical communities. We used artificial intelligence models dubbed transformers to generate false cybersecurity news and COVID-19 medical studies and presented the cybersecurity misinformation to cybersecurity experts for testing. We found that transformer-generated misinformation was able to fool cybersecurity experts.</p>
<h2>Transformers</h2>
<p>Much of the technology used to identify and manage misinformation is powered by artificial intelligence. AI allows computer scientists to fact-check large amounts of misinformation quickly, given that there’s too much for people to detect without the help of technology. Although AI helps people detect misinformation, it has ironically also been used to produce misinformation in recent years. </p>
<figure class="align-right zoomable">
<a href="https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="A block of text on a smartphone screen" src="https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=237&fit=clip" srcset="https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=355&fit=crop&dpr=1 600w, https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=355&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=355&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=447&fit=crop&dpr=1 754w, https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=447&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/404378/original/file-20210603-19-13kxhs4.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=447&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">AI can help detect misinformation like these false claims about COVID-19 in India – but what happens when AI is used to generate the misinformation?</span>
<span class="attribution"><a class="source" href="https://newsroom.ap.org/detail/VirusOutbreakMisinformationIndia/d455fd7187004eb9a65472675ee4b3b4/photo">AP Photo/Ashwini Bhatia</a></span>
</figcaption>
</figure>
<p>Transformers, like <a href="https://searchengineland.com/welcome-bert-google-artificial-intelligence-for-understanding-search-queries-323976">BERT</a> from Google and <a href="https://openai.com/blog/better-language-models/">GPT</a> from OpenAI, use <a href="https://www.cio.com/article/3258837/natural-language-processing-nlp-explained.html">natural language processing</a> to understand text and produce translations, summaries and interpretations. They have been used in such tasks as storytelling and answering questions, pushing the boundaries of machines displaying humanlike capabilities in generating text.</p>
<p>Transformers have aided Google and other technology companies by <a href="https://blog.google/products/search/search-language-understanding-bert/">improving their search engines</a> and have helped the general public in combating such common problems as <a href="https://www.newyorker.com/culture/cultural-comment/the-computers-are-getting-better-at-writing">battling writer’s block</a>.</p>
<p>Transformers can also be used for malevolent purposes. Social networks like Facebook and Twitter have already faced the challenges of <a href="https://www.technologyreview.com/2020/01/08/130983/were-fighting-fake-news-ai-bots-by-using-more-ai-thats-a-mistake/">AI-generated fake news</a> across platforms.</p>
<h2>Critical misinformation</h2>
<p>Our research shows that transformers also pose a misinformation threat in medicine and cybersecurity. To illustrate how serious this is, we <a href="https://ruder.io/recent-advances-lm-fine-tuning/">fine-tuned</a> the GPT-2 transformer model on <a href="https://www.cisecurity.org/blog/what-is-cyber-threat-intelligence/">open online sources</a> discussing cybersecurity vulnerabilities and attack information. A cybersecurity vulnerability is the weakness of a computer system, and a cybersecurity attack is an act that exploits a vulnerability. For example, if a vulnerability is a weak Facebook password, an attack exploiting it would be a hacker figuring out your password and breaking into your account. </p>
<p>We then seeded the model with the sentence or phrase of an actual cyberthreat intelligence sample and had it generate the rest of the threat description. We presented this generated description to cyberthreat hunters, who sift through lots of information about cybersecurity threats. These professionals read the threat descriptions to identify potential attacks and adjust the defenses of their systems. </p>
<p>We were surprised by the results. The cybersecurity misinformation examples we generated were able to fool cyberthreat hunters, who are knowledgeable about all kinds of cybersecurity attacks and vulnerabilities. Imagine this scenario with a crucial piece of cyberthreat intelligence that involves the airline industry, which we generated in our study.</p>
<figure class="align-center ">
<img alt="A block of text with false information about a cybersecurity attack on airlines" src="https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=162&fit=crop&dpr=1 600w, https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=162&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=162&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=203&fit=crop&dpr=1 754w, https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=203&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/404375/original/file-20210603-15-y1w385.JPG?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=203&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">An example of AI-generated cybersecurity misinformation.</span>
<span class="attribution"><span class="source">The Conversation</span>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span>
</figcaption>
</figure>
<p>This misleading piece of information contains incorrect information concerning cyberattacks on airlines with sensitive real-time flight data. This false information could keep cyber analysts from addressing legitimate vulnerabilities in their systems by shifting their attention to fake software bugs. If a cyber analyst acts on the fake information in a real-world scenario, the airline in question could have faced a serious attack that exploits a real, unaddressed vulnerability.</p>
<p>A similar transformer-based model can generate information in the medical domain and potentially fool medical experts. During the COVID-19 pandemic, preprints of research papers that have not yet undergone a rigorous review are constantly being uploaded to such sites as <a href="https://www.medrxiv.org/">medrXiv</a>. They are not only being described in the press but are being used to make public health decisions. Consider the following, which is not real but generated by our model after minimal fine-tuning of the default GPT-2 on some COVID-19-related papers.</p>
<figure class="align-center ">
<img alt="A block of text showing health care misinformation." src="https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=282&fit=crop&dpr=1 600w, https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=282&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=282&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=355&fit=crop&dpr=1 754w, https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=355&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/404376/original/file-20210603-21-1ool1co.JPG?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=355&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">An example of AI-generated health care misinformation.</span>
<span class="attribution"><span class="source">The Conversation</span>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span>
</figcaption>
</figure>
<p>The model was able to generate complete sentences and form an abstract allegedly describing the side effects of COVID-19 vaccinations and the experiments that were conducted. This is troubling both for medical researchers, who consistently rely on accurate information to make informed decisions, and for members of the general public, who often rely on public news to learn about critical health information. If accepted as accurate, this kind of misinformation could put lives at risk by misdirecting the efforts of scientists conducting biomedical research.</p>
<p>[<em><a href="https://theconversation.com/us/newsletters/science-editors-picks-71/?utm_source=TCUS&utm_medium=inline-link&utm_campaign=newsletter-text&utm_content=science-corona-important">The Conversation’s most important coronavirus headlines, weekly in a science newsletter</a></em>]</p>
<h2>An AI misinformation arms race?</h2>
<p>Although examples like these from our study can be fact-checked, transformer-generated misinformation hinders such industries as health care and cybersecurity in adopting AI to help with information overload. For example, automated systems are being developed to extract data from cyberthreat intelligence that is then used to inform and train automated systems to recognize possible attacks. If these automated systems process such false cybersecurity text, they will be less effective at detecting true threats.</p>
<p>We believe the result could be an arms race as people spreading misinformation develop better ways to create false information in response to effective ways to recognize it.</p>
<p>Cybersecurity researchers continuously study ways to detect misinformation in different domains. Understanding how to automatically generate misinformation helps in understanding how to recognize it. For example, automatically generated information often has subtle grammatical mistakes that systems can be trained to detect. Systems can also cross-correlate information from multiple sources and identify claims lacking substantial support from other sources. </p>
<p>Ultimately, everyone should be more vigilant about what information is trustworthy and be aware that hackers exploit people’s credulity, especially if the information is not from reputable news sources or published scientific work.</p><img src="https://counter.theconversation.com/content/160909/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Priyanka Ranade is a Computer Science PhD Student at the University of Maryland, Baltimore County. In addition to her role at UMBC, she is affiliated with Northrop Grumman Corporation and is an Adjunct Professor at the University of Maryland, College Park.
</span></em></p><p class="fine-print"><em><span>Anupam Joshi receives funding from a variety of federal (e.g. NSF, DoD, NIST) and industry (e.g. Google, IBM) sources for his research and education related activities. In addition to his primary appointment at UMBC, he is an Affiliate Professor in the School of Medicine at the University of Maryland Baltimore, and an Adjunct Professor at IIT Delhi. </span></em></p><p class="fine-print"><em><span>Tim Finin receives funding from both federal agencies (e.g., NSF, DoD, NIST) and industry (e.g., Google, IBM) to support his research. In addition to his primary appointment at UMBC, he is affiliated with the Human Language Technology Center of Excellence at Johns Hopkins University.</span></em></p>Bots flooding social media with fake news about politics is bad enough. Muddying the waters in such fields as cybersecurity and health care could put lives at risk.Priyanka Ranade, PhD Student in Computer Science and Electrical Engineering, University of Maryland, Baltimore CountyAnupam Joshi, Professor of Computer Science & Electrical Engineering, University of Maryland, Baltimore CountyTim Finin, Professor of Computer Science and Electrical Engineering, University of Maryland, Baltimore CountyLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1594302021-05-31T20:09:35Z2021-05-31T20:09:35ZMachine learning is changing our culture. Try this text-altering tool to see how<figure><img src="https://images.theconversation.com/files/403503/original/file-20210531-19-hh9v4l.jpg?ixlib=rb-1.1.0&rect=125%2C71%2C5865%2C3916&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><span class="source">Shutterstock</span></span></figcaption></figure><p>Most of us benefit every day from the fact computers can now “understand” us when we speak or write. Yet few of us have paused to consider the potentially damaging ways this same technology may be shaping our culture. </p>
<p>Human language is full of ambiguity and double meanings. For instance, consider the potential meaning of this phrase: “I went to project class”. Without context, it’s an ambiguous statement.</p>
<p>Computer scientists and linguists have spent decades trying to program computers to understand the nuances of human language. And in certain ways, computers are fast approaching humans’ ability to understand and <a href="https://culturalanalytics.org/article/17212.pdf">generate text</a>.</p>
<p>Through the very act of suggesting some words and not others, the predictive text and auto-complete features in our devices change the way we think. Through these subtle, everyday interactions, machine learning is influencing our culture. Are we ready for that?</p>
<p>I created an online interactive work for the <a href="https://www.kyoglewritersfestival.com/">Kyogle Writers Festival</a> that lets you explore this technology in a harmless way.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=450&fit=crop&dpr=1 600w, https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=450&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=450&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=566&fit=crop&dpr=1 754w, https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=566&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/403493/original/file-20210531-13-1wzkdxj.JPG?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=566&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">A photo from the Kyogle Writers Festival in NSW, earlier this year.</span>
<span class="attribution"><span class="license">Author provided</span></span>
</figcaption>
</figure>
<h2>What is natural language processing?</h2>
<p>The field concerned with using everyday language to interact with computers is called “natural language processing”. We encounter it when we speak to Siri or Alexa, or type words into a browser and have the rest of our sentence predicted.</p>
<p>This is only possible due to vast improvements in natural language processing over the past decade — achieved through sophisticated machine-learning algorithms trained on enormous datasets (usually billions of words).</p>
<p>Last year, this technology’s potential became clear when the Generative Pre-trained Transformer 3 (GPT-3) was released. It set a new benchmark in what computers can do with language. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/can-robots-write-machine-learning-produces-dazzling-results-but-some-assembly-is-still-required-146090">Can robots write? Machine learning produces dazzling results, but some assembly is still required</a>
</strong>
</em>
</p>
<hr>
<p>GPT-3 can take just a few words or phrases and generate whole documents <a href="https://www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-article-gpt-3">of “meaningful” language</a>, by capturing the contextual relationships between words in a sentence. It does this by building on machine-learning models, including two widely adopted models called <a href="http://jalammar.github.io/illustrated-bert/">“BERT” and “ELMO”</a>.</p>
<h2>How is this technology affecting culture?</h2>
<p>However, there is a key issue with any language model produced by machine learning: they generally learn everything they know from data sources such as Wikipedia and Twitter.</p>
<p>In effect, machine learning takes data from the past, “learns” from it to produce a model, and uses this model to carry out tasks in the future. But during this process, <a href="https://dl.acm.org/doi/pdf/10.1145/3442188.3445922">a model</a> may absorb a distorted or problematic worldview from its training data.</p>
<p>If the training data was biased, this bias will be codified and reinforced in the model, rather than being challenged. For example, a model may end up associating certain identity groups or races with positive words, and others with negative words. </p>
<p>This can lead to serious exclusion and inequality, as detailed in the recent documentary <a href="https://www.ajl.org/">Coded Bias</a>. </p>
<h2>Everything you ever said</h2>
<p>The interactive work I created allows people to playfully gain an intuition for how computers understand language. It is called Everything You Ever Said (EYES), in reference to the way natural language models draw on all kinds of data sources for training.</p>
<p>EYES allows you to take any piece of writing (less than 2000 characters) and “subtract” one concept and “add” another. In other words, it lets you use a computer to change the meaning of a piece of text. You can <a href="https://www.everythingyoueversaid.art">try it yourself</a>.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&rect=0%2C275%2C1908%2C813&q=45&auto=format&w=1000&fit=clip"><img alt="Screenshot of natural language processing tool" src="https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&rect=0%2C275%2C1908%2C813&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=319&fit=crop&dpr=1 600w, https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=319&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=319&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=401&fit=crop&dpr=1 754w, https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=401&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/403499/original/file-20210531-15-1e10z30.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=401&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">EYES can add and subtract concepts from the text you input, based on an understanding of English from training data.</span>
<span class="attribution"><a class="source" href="https://www.everythingyoueversaid.art/">Screenshot</a></span>
</figcaption>
</figure>
<p>Here’s an example of the Australian national anthem subjected to some automated revision. I subtracted the concept of “empire” and added the concept of “koala” to get:</p>
<p><em>Australians all let us grieve<br>
For we are one and free<br>
We’ve golden biota and abundance for poorness<br>
Our koala is girt by porpoise<br>
Our wildlife abounds in primate’s koalas<br>
Of naturalness shiftless and rare<br>
In primate’s wombat, let every koala<br>
Wombat koala fair<br>
In joyous aspergillosis then let us vocalise,<br>
Wombat koala fair</em></p>
<p>What is going on here? At its core, EYES uses a model of the English language developed by researchers from Stanford University in the United States, called <a href="https://nlp.stanford.edu/projects/glove/">GLoVe</a> (Global Vectors for Word Representation).</p>
<p>EYES uses GLoVe to change the text by making a series of analogies, wherein an “analogy” is a comparison between one thing and another. For instance, if I ask you: “man is to king what woman is to?” — you might answer “queen”. That’s an easy one.</p>
<p>But I could ask a more challenging question such as: “rose is to thorn what love is to?” There are several possible answers here, depending on your interpretation of the language. When asked about these analogies, GLoVe will produce the responses “queen” and “betrayal”, respectively. </p>
<p>GLoVe has every word in the English language represented as a vector in a multi-dimensional space (of around 300 dimensions). A such, it can perform calculations with words, adding and subtracting words as if they were numbers. </p>
<h2>Cyborg culture is already here</h2>
<p>The trouble with machine learning is that the associations being made between certain concepts remain hidden inside a black box; we can’t see or touch them. Approaches to making machine learning models more transparent are a <a href="https://www.scientificamerican.com/article/demystifying-the-black-box-that-is-ai/">focus of much current research</a>. </p>
<p>The purpose of EYES is to let you experiment with these associations in a more playful way, so you can develop an intuition for how machine learning models view the world. </p>
<p>Some analogies will surprise you with their poignancy, while others may well leave you bewildered. Yet, every association was inferred from a huge corpus of a few billion words written by ordinary people.</p>
<p>Models such as GPT-3, which have learned from similar data sources, are already influencing how we use language. Having entire news feeds populated by machine-written text is no longer the stuff of science fiction. This technology is <a href="https://notrealnews.net/">already here</a>. </p>
<p>And the cultural footprint of machine-learning models seems to only be growing. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/gpt-3-new-ai-can-write-like-a-human-but-dont-mistake-that-for-thinking-neuroscientist-146082">GPT-3: new AI can write like a human but don't mistake that for thinking – neuroscientist</a>
</strong>
</em>
</p>
<hr>
<img src="https://counter.theconversation.com/content/159430/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Nick Kelly has received past funding from the Queensland State Government and the Queensland College of Teachers. He would like to thank the 2021 Kyogle Writers Festival for including the work in its program and Jess Greentree and Sam Hobson for their feedback on early drafts of the work.</span></em></p>Through the act of suggesting some words and not others, the predictive text features in our devices change the way we think — and therefore shape our culture.Nick Kelly, Senior Lecturer in Interaction Design, Queensland University of TechnologyLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1596602021-04-28T13:06:17Z2021-04-28T13:06:17Z‘Can I see your parts list?’ What AI’s attempted chat-up lines tell us about computer-generated language<figure><img src="https://images.theconversation.com/files/397566/original/file-20210428-23-icwlyq.png?ixlib=rb-1.1.0&rect=181%2C191%2C1875%2C1256&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><a class="source" href="https://www.shutterstock.com/image-vector/usb-flash-heart-connection-105676262">Sararoom Design/Shutterstock</a></span></figcaption></figure><p>Have you ever wondered what flirting with artificial intelligence would look like? Research scientist and engineer Janelle Shane has given us an idea by training a <a href="https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414">neural network</a> – an algorithm loosely inspired by biological brain structures – to produce <a href="https://janellecshane.substack.com/p/gpt-3-tries-pickup-lines?utm_source=Nature+Briefing&utm_campaign=686d1220e7-briefing-dy-20210407&utm_medium=email&utm_term=0_c9dfd39373-686d1220e7-46275490">chat-up lines</a>. </p>
<p>Some of the results are hilarious and completely nonsensical, such as the inelegant: “2017 Rugboat 2-tone Neck Tie Shirt”. But some of them turned out pretty well. At least, if you’re a robot:</p>
<blockquote>
<p>I can tell by your red power light that you’re into me.</p>
<p>You look like a thing and I love you.</p>
<p>Can I see your parts list?</p>
</blockquote>
<p>But how were these lines generated, and why do the results vary so much in terms of quality and cohesiveness? That’s down to the types of neural networks Shane worked with: all based on <a href="https://www.nytimes.com/2020/11/24/science/artificial-intelligence-ai-gpt3.html">GPT-3</a>, the world’s largest language model to date. </p>
<h2>Language modelling</h2>
<p>GPT stands for generative pre-trained transformer. Its current version, developed by <a href="https://openai.com/">OpenAI</a>, is the third in a line of ever-improving natural language processing systems trained to produce human-like text or speech.</p>
<p><a href="https://towardsdatascience.com/introduction-to-natural-language-processing-nlp-323cc007df3d">Natural language processing</a>, or NLP, refers to the application of computers to process and generate large amounts of coherent spoken or written text. Whether you ask Siri for a weather update, request for Alexa to turn on the lights, or you use Google to translate a message from French into English, you’re able to do so because of developments in NLP. </p>
<figure class="align-center ">
<img alt="A Google Home speaker photographed from above." src="https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=451&fit=crop&dpr=1 600w, https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=451&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=451&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=566&fit=crop&dpr=1 754w, https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=566&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/397571/original/file-20210428-23-1vv07j.jpeg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=566&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Voice-activated smart speakers use NLP technology to understand your spoken requests.</span>
<span class="attribution"><a class="source" href="https://www.shutterstock.com/image-photo/smart-home-speaker-device-voice-activated-1537574132">Vantage_DS/Shutterstock</a></span>
</figcaption>
</figure>
<p>It takes a variety of NLP tasks – from speech recognition to picking apart sentence structures – for applications such as Siri to successfully requests. The virtual assistant, much like any other language-based tool, is trained using many thousands of sentences, ideally as varied and diverse as possible. </p>
<p>Because human language is extremely complex, the best NLP applications rely increasingly on pre-trained models that allow “<a href="https://www.aclweb.org/anthology/E17-1096/">contextual bidirectional learning</a>”. This means considering a word’s wider context in a sentence, scanning both left and right of any given word to identify the word’s intended meaning. More recent models can even pay attention to more nuanced features of human language, <a href="https://arxiv.org/abs/2011.11465">such as irony and sarcasm</a>.</p>
<h2>Computer compliments</h2>
<p>GPT-3 is such a successful language-generating AI because it doesn’t need retraining over and over again to complete a new task. Instead, it uses what the model has already learned about language and applies it to a <a href="https://openai.com/blog/gpt-3-apps/">something new</a> – such as <a href="https://www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-article-gpt-3">writing articles</a> and <a href="https://www.wired.com/story/ai-latest-trick-writing-computer-code/">computer code</a>, generating <a href="https://www.inputmag.com/gaming/sorry-video-game-developers-ai-is-coming-for-your-job">novel dialogue in video games</a>, or formulating chat-up lines.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/robo-journalism-computer-generated-stories-may-be-inevitable-but-its-not-all-bad-news-89473">Robo-journalism: computer-generated stories may be inevitable, but it's not all bad news</a>
</strong>
</em>
</p>
<hr>
<p>Compared to its predecessor GPT-2, the third-generation model is 116 times bigger and has been trained on billions of words of data. To generate its chat-up lines, GPT-3 was simply asked to automate the text for an article headlined: “These are the top pickup lines of 2021! Amaze your crush and get results!”</p>
<p>Because GPT-3’s training updates have been added gradually over time, this same prompt could also be used on smaller, more basic variants – generating weirder and less coherent chat-up lines:</p>
<blockquote>
<p>Hey, my name is John Smith. Will you sit on my breadbox while I cook or is there some kind of speed limit on that thing?</p>
<p>It is urgent that you become a professional athlete.</p>
<p>CAPE FASHION</p>
</blockquote>
<p>But GPT-3’s “DaVinci” variant – its largest and most competent iteration to date – delivered some more convincing attempts which might actually pass for effective flirting – with a little fine-tuning:</p>
<blockquote>
<p>You have the most beautiful fangs I’ve ever seen.</p>
<p>I love you. I don’t care if you’re a doggo in a trenchcoat.</p>
<p>I have exactly 4 stickers. I need you to be the 5th.</p>
</blockquote>
<p>The latest variant of GPT-3 is currently the largest contextual language model in the world and is able to complete a number of highly impressive tasks. But is it smart enough to pass as a human? </p>
<h2>Almost human</h2>
<p>As one of the pioneers of modern computing and a firm believer in true artificial intelligence, Alan Turing developed the “Imitation Game” in 1950 – today known as the “<a href="https://plato.stanford.edu/entries/turing-test/">Turing Test</a>”. If a computer’s performance is indistinguishable from that of a human, it passes the Turing Test. In language generation alone, GPT-3 could soon pass Alan Turing’s test.</p>
<p>But it doesn’t really matter if GPT-3 passes the Turing Test or not. Its performance is likely to depend on the specific task the model is used for – which, judging by the technology’s flirting, should probably be something other than the delicate art of the chat-up line.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/gpt-3-new-ai-can-write-like-a-human-but-dont-mistake-that-for-thinking-neuroscientist-146082">GPT-3: new AI can write like a human but don't mistake that for thinking – neuroscientist</a>
</strong>
</em>
</p>
<hr>
<p>And, even if it were to pass the Turing Test, in no way would this make the model truly intelligent. At best, it would be extremely well trained on <a href="https://link.springer.com/article/10.1007/s11023-020-09548-1">specific semantic tasks</a>. Maybe the more important question to ask is: do we even want to make GPT-3 more human? </p>
<h2>Learning from humans</h2>
<p>Shortly after its reveal in summer 2020, GPT-3 <a href="https://www.nytimes.com/2020/11/24/science/artificial-intelligence-ai-gpt3.html">made headlines</a> for spewing out shockingly sexist and racist content. But this was hardly surprising. The language generator was trained on vast amounts of text on the internet, and without remodelling and <a href="https://theconversation.com/online-translators-are-sexist-heres-how-we-gave-them-a-little-gender-sensitivity-training-157846">retraining</a> it was doomed to replicate the <a href="https://thenextweb.com/news/gpt-3-is-the-worlds-most-powerful-bigotry-generator-what-should-we-do-about-it">biases</a>, harmful language and misinformation that we know to exist online.</p>
<p>Clearly, language models such as GPT-3 do not come without <a href="https://www.bbc.co.uk/news/technology-49446729">potential risks</a>. If we want these systems to be the basis of our digital assistants or conversational agents, we need to be more rigorous and selective when giving them reading material to learn from. </p>
<p>Still, <a href="https://arxiv.org/abs/2103.12407">recent research</a> has shown that GPT-3’s knowledge of the internet’s dark side could actually be used to automatically detect online hate speech, with up to 78% accuracy. So even though its chat-up lines look unlikely to kindle more love in the world, GPT-3 could may be set, at least, to reduce the hate.</p><img src="https://counter.theconversation.com/content/159660/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Stefanie Ullmann is affiliated with the Giving Voice to Digital Democracies project, funded by the International Foundation for the Humanities and Social Change.</span></em></p>Endearing and amusing, AI’s faltering attempts at flirting show how far computer-generated language has come.Stefanie Ullmann, Postdoctoral Research Associate, Centre for the Humanities and Social Change, University of CambridgeLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1473852020-10-19T11:43:39Z2020-10-19T11:43:39ZTeaching computers to read health records is helping fight COVID-19 – here’s how<p>Medical records are a rich source of health data. When combined, the information they contain can help researchers <a href="https://www.bhf.org.uk/%7E/media/files/publications/research/clear-and-present-data.pdf#page=4">better understand diseases</a> and treat them more effectively. This includes COVID-19. But to unlock this rich resource, researchers first need to read it.</p>
<p>We may have moved on from the days of handwritten medical notes, but the information recorded in modern electronic health records can be just as hard to access and interpret. It’s an old joke that doctors’ handwriting is illegible, but it turns out their typing isn’t much better. </p>
<p>The sheer volume of information contained in health records is staggering. Every day, healthcare staff in a typical NHS hospital generate so much text it would take a human an age just to scroll through it, let alone read it. Using computers to analyse all this data is an obvious solution, but far from simple. What makes perfect sense to a human can be highly difficult for a computer to understand.</p>
<p><a href="https://www.hdruk.ac.uk/projects/national-text-analytics-project/">Our team</a> is using a form artificial intelligence to bridge this gap. By teaching computers how to comprehend human doctors’ notes, we’re hoping they’ll uncover insights on how to fight COVID-19 by finding patterns across many thousands of patients’ records.</p>
<h2>Why health records are hard going</h2>
<p>A significant proportion of a health record is made up of free text, typed in narrative form like an email. This includes the patient’s symptoms, the history of their illness, and notes about pre-existing conditions and medications they’re taking. There may also be relevant information about family members and lifestyle mixed in too. And because this text has been entered by busy doctors, there will also be abbreviations, inaccuracies and typos. </p>
<figure class="align-center ">
<img alt="A doctor using a computer" src="https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=450&fit=crop&dpr=1 600w, https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=450&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=450&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=566&fit=crop&dpr=1 754w, https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=566&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/362700/original/file-20201009-13-1as206v.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=566&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Doctors write information in free text boxes is rich in detail but poorly arranged for a machine to understand.</span>
<span class="attribution"><a class="source" href="https://www.shutterstock.com/image-photo/doctor-using-computer-portrait-5902651">logoboom/Shutterstock</a></span>
</figcaption>
</figure>
<p>This kind of information is known as unstructured data. For example, a patient’s record might say:</p>
<blockquote>
<p>Mrs Smith is a 65-year-old woman with atrial fibrillation and had a CVA in March. She had a past history of a #NOF and OA. Family history of breast cancer. She has been prescribed apixaban. No history of haemorrhage. </p>
</blockquote>
<p>This highly compact paragraph contains a large amount of data about Mrs Smith. Another human reading the notes would know what information is important and be able to extract it in seconds, but a computer would find the task extremely difficult.</p>
<h2>Teaching machines to read</h2>
<p>To solve this problem, we’re using something called <a href="https://towardsdatascience.com/your-guide-to-natural-language-processing-nlp-48ea2511f6e1">natural language processing</a> (NLP). Based on machine learning and AI technology, NLP algorithms translate the language used in free text into a standardised, structured set of medical terms that can be analysed by a computer. </p>
<p>These algorithms are extremely complex. They need to understand context, long strings of words and medical concepts, distinguish current events from historic ones, identify family relationships and more. We teach them to do this by feeding them existing written information so they can learn the structure and meaning of language – in this case, publicly available English text from the internet – and then use real medical records for further improvement and testing. </p>
<p>Using NLP algorithms to analyse and extract data from health records has huge potential to change healthcare. Much of what’s captured in narrative text in a patient’s notes is normally never seen again. This could be important information such as the early warning signs of serious diseases like cancer or stroke. Being able to automatically analyse and flag important issues could help deliver better care and avoid delays in diagnosis and treatment.</p>
<h2>Finding ways to fight COVID-19</h2>
<p>By drawing together health records using these tools, we’re now using these techniques to see patterns that are relevant to the pandemic. For example, we recently used our tools to discover whether drugs commonly prescribed to treat high blood pressure, diabetes and other conditions – known as <a href="https://www.bhf.org.uk/informationsupport/heart-matters-magazine/medical/drug-cabinet/ace-inhibitors">angiotensin-converting enzyme inhibitors (ACEIs)</a> and <a href="https://www.bhf.org.uk/informationsupport/heart-matters-magazine/medical/drug-cabinet/arbs">angiotensin receptor blockers (ARBs)</a> – increase the chances of becoming severely ill with COVID-19.</p>
<p>The virus that causes COVID-19 infects cells by binding to a molecule on the cell surface called <a href="https://theconversation.com/ace2-the-molecule-that-helps-coronavirus-invade-your-cells-138369">ACE2</a>. Both ACEIs and ARBs are thought to <a href="https://www.nice.org.uk/advice/es24/chapter/Key-messages">increase the amount of ACE2</a> on the surface of cells, leading to concerns that these drugs could be putting people at increased risk from the virus. </p>
<figure class="align-center ">
<img alt="SARS-CoV-2 binding with ACE2 on the surface of a cell" src="https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=300&fit=crop&dpr=1 600w, https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=300&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=300&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=377&fit=crop&dpr=1 754w, https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=377&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/362508/original/file-20201008-24-1d2fl23.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=377&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">The coronavirus (red) binds with ACE2 proteins (blue) on the cell’s surface (green) to gain entry.</span>
<span class="attribution"><a class="source" href="https://www.shutterstock.com/image-illustration/sarscov2-viruses-binding-ace2-receptors-on-1701661810">Kateryna Kon/Shutterstock</a></span>
</figcaption>
</figure>
<p>However, the information needed to answer this question – how many severely ill COVID-19 patients are being prescribed these drugs – can be recorded both as structured prescriptions and in free text in their medical records. That free text needs to be in a computer-searchable format for a machine to answer the question.</p>
<p>Using our NLP tools, we were able to analyse the anonymised records of 1,200 COVID-19 patients, comparing clinical outcomes with whether or not patients were taking these drugs. Reassuringly, we found that people prescribed ACEIs or ARBs were <a href="https://doi.org/10.1002/ejhf.1924">no more likely</a> to be severely ill than those not taking the drugs.</p>
<p>We’re now expanding how we use these tools to find out more about who is most at risk from COVID-19. For instance, we’ve used them to <a href="https://www.thelancet.com/journals/eclinm/article/PIIS2589-5370(20)30318-7/fulltext">investigate the links</a> between ethnicity, pre-existing health conditions and COVID-19. This has revealed several striking things: that being black or of mixed ethnicity makes you more likely to be admitted to hospital with the disease, and that Asian patients, when in hospital, are at greater risk of being admitted to intensive care or dying from COVID-19.</p>
<p>We’ve also used these tools to evaluate the <a href="https://www.medrxiv.org/content/10.1101/2020.04.24.20078006v3">early warning scores</a> that predict which patients admitted to hospital are most likely to become severely ill, and to suggest what additional measures could be used to improve these scores. We’re also using the technology to <a href="https://www.medrxiv.org/content/10.1101/2020.10.02.20205617v1">predict upcoming surges</a> of COVID-19 cases, based on patients’ symptoms that doctors have recorded.</p><img src="https://counter.theconversation.com/content/147385/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>James Teo received has research support from Innovate UK, the UK government's Office of Life Sciences, Bristol Myers Squibb, the NIHR Applied Research Centre South London, London Medical Imaging and AI Centre for Value-Based Healthcare (AI4VBH) and the Health Innovation Network.</span></em></p><p class="fine-print"><em><span>Richard Dobson has received funding from the Motor Neurone Disease Association, the Maudsley Charity, MND Scotland, Innovate UK, Takeda California Inc., the European Commission, Health Data Research UK, the Medical Research Council, the Psychiatry Research Trust, the National Institute for Health Research, Alzheimer's Research UK, Guy's and St Thomas' Charity, Janssen Pharmaceutica N.V., Ochre Bio Ltd and Glaxo Wellcome Research & Development Ltd.</span></em></p>Decoding doctors’ writing can unlock vital health data.James Teo, Neurologist, Clinical Director of Data and AI and Clinical Senior Lecturer, King's College LondonRichard Dobson, Professor in Health Informatics, King's College LondonLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1321582020-03-09T12:22:22Z2020-03-09T12:22:22ZHow technology can combat the rising tide of fake science<figure><img src="https://images.theconversation.com/files/318753/original/file-20200304-66112-vybpt.jpg?ixlib=rb-1.1.0&rect=63%2C13%2C1178%2C840&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">A crop circle in Switzerland.</span> <span class="attribution"><a class="source" href="https://commons.wikimedia.org/wiki/File:CropCircleW.jpg">Jabberocky/Wikimedia Commons</a></span></figcaption></figure><p>Science gets a lot of respect these days. Unfortunately, it’s also getting a lot of competition from misinformation. Seven in 10 Americans think the benefits from science outweigh the harms, and nine in 10 think science and technology will create <a href="https://nsf.gov/statistics/2018/nsb20181/report/sections/science-and-technology-public-attitudes-and-understanding/highlights">more opportunities for future generations</a>. Scientists have made dramatic progress in understanding the universe and the mechanisms of biology, and advances in computation benefit all fields of science. </p>
<p>On the other hand, Americans are surrounded by a rising tide of misinformation and fake science. Take climate change. Scientists are in <a href="https://doi.org/10.1088/1748-9326/8/2/024024">almost complete agreement that people are the primary cause of global warming</a>. Yet polls show that <a href="https://doi.org/10.1007/s10584-019-02406-9">a third of the public disagrees</a> with this conclusion.</p>
<p>In my <a href="https://scholar.google.com/citations?user=OrRLRQ4AAAAJ&hl=en&oi=ao">30 years of studying and promoting scientific literacy</a>, I’ve found that college educated adults have large holes in their basic science knowledge and they’re disconcertingly <a href="https://ejse.southwestern.edu/article/view/17315">susceptible to superstition and beliefs that aren’t based on any evidence</a>. One way to counter this is to make it easier for people to detect pseudoscience online. To this end, my lab at the University of Arizona has developed an artificial intelligence-based pseudoscience detector that we plan to freely release as a web browser extension and smart phone app.</p>
<h2>Americans’ predilection for fake science</h2>
<p>Americans are prone to superstition and paranormal beliefs. An annual survey done by sociologists at Chapman University finds that <a href="https://blogs.chapman.edu/wilkinson/2018/10/16/paranormal-america-2018/">more than half believe in spirits and the existence of ancient civilizations</a> like Atlantis, and more than a third think that aliens have visited the Earth in the past or are visiting now. Over 75% hold multiple paranormal beliefs. The survey shows that these numbers have increased in recent years.</p>
<p><iframe id="IbP7D" class="tc-infographic-datawrapper" src="https://datawrapper.dwcdn.net/IbP7D/3/" height="400px" width="100%" style="border: none" frameborder="0"></iframe></p>
<p>Widespread belief in astrology is a pet peeve of my colleagues in astronomy. It’s long had a foothold in the popular culture through horoscopes in newspapers and magazines <a href="https://www.theatlantic.com/health/archive/2018/01/the-new-age-of-astrology/550034/">but currently it’s booming</a>. Belief is strong even among the most educated. My surveys of college undergraduates show that three-quarters of them <a href="http://dx.doi.org/10.3847/AER2010040">think that astrology is very or “sort of” scientific</a> and only half of science majors recognize it as not at all scientific.</p>
<p>Allan Mazur, a sociologist at Syracuse University, has delved into <a href="https://www.taylorfrancis.com/books/9780203788967">the nature of irrational belief systems</a>, their cultural roots, and their political impact. Conspiracy theories are, by definition, resistant to evidence or data that might prove them false. Some are at least amusing. Adherents of the flat Earth theory turn back the clock on two millennia of scientific progress. <a href="https://www.theverge.com/2017/10/9/16424622/reddit-conspiracy-theories-memes-irony-flat-earth">Interest in this bizarre idea has surged in the past five years</a>, spurred by social media influencers and the echo chamber nature of web sites like Reddit. As with climate change denial, <a href="https://www.bbc.com/news/technology-47279253">many come to this belief through YouTube videos</a>.</p>
<p>However, the consequences of fake science are no laughing matter. In matters of health and climate change, <a href="https://doi.org/10.1098/rsos.190161">misinformation can be a matter of life and death</a>. Over a 90-day period spanning December, January and February, people liked, shared and commented on posts from sites containing <a href="https://www.zdnet.com/article/coronavirus-misinformation-is-increasing-newsguard-finds/">false or misleading information about COVID-19</a> 142 times more than they did information from the Centers for Disease Control and the World Health Organization. </p>
<p>Combating fake science is an urgent priority. In a world that’s increasingly dependent on science and technology, civic society can only function when the electorate is well informed. </p>
<p>Educators must roll up their sleeves and do a better job of teaching critical thinking to young people. However, the problem goes beyond the classroom. The internet is the <a href="https://www.nsf.gov/statistics/2018/nsb20181/report">first source of science information</a> for 80% of people ages 18 to 24. </p>
<p>One study found that a majority of a random sample of 200 YouTube videos on climate change <a href="https://doi.org/10.3389/fcomm.2019.00036">denied that humans were responsible or claimed that it was a conspiracy</a>. The videos peddling conspiracy theories got the most views. Another study found that <a href="https://www.theguardian.com/technology/2020/feb/21/climate-tweets-twitter-bots-analysis">a quarter of all tweets on climate were generated by bots</a> and they preferentially amplified messages from climate change deniers.</p>
<h2>Technology to the rescue?</h2>
<p>The recent success of machine learning and AI in <a href="https://arxiv.org/abs/1705.00648">detecting fake news</a> points the way to detecting fake science online. The key is <a href="https://www.explainthatstuff.com/introduction-to-neural-networks.html">neural net</a> technology. Neural nets are loosely modeled on the human brain. They consist of many interconnected computer processors that identify meaningful patterns in data like words and images. Neural nets already permeate everyday life, particularly in <a href="https://arxiv.org/abs/1708.02709">natural language processing</a> systems like Amazon’s Alexa and Google’s language translation capability.</p>
<p>At the University of Arizona, we have trained neural nets on handpicked popular articles about climate change and biological evolution, and the neural nets are 90% successful in distinguishing wheat from chaff. With a quick scan of a site, our neural net can tell if its content is scientifically sound or climate-denial junk. After more refinement and testing we hope to have neural nets that can work across all domains of science. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=471&fit=crop&dpr=1 600w, https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=471&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=471&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=591&fit=crop&dpr=1 754w, https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=591&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/318416/original/file-20200303-66064-2dk56c.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=591&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Neural net technology under development at the University of Arizona will flag science websites with a color code indicating their reliability (left). A smartphone app version will gamify the process of declaring science articles real or fake (right).</span>
<span class="attribution"><span class="source">Chris Impey</span>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span>
</figcaption>
</figure>
<p>The goal is a web browser extension that would detect when the user is looking at science content and deduce whether or not it’s real or fake. If it’s misinformation, the tool will suggest a reliable web site on that topic. My colleagues and I also plan to gamify the interface with a smart phone app that will let people compete with their friends and relatives to detect fake science. Data from the best of these participants will be used to help train the neural net.</p>
<p>Sniffing out fake science should be easier than sniffing out fake news in general, because subjective opinion plays a minimal role in legitimate science, which is characterized by evidence, logic and verification. Experts can readily distinguish legitimate science from conspiracy theories and arguments motivated by ideology, which means machine learning systems can be trained to, as well. </p>
<p>“Everyone is entitled to his own opinion, but not his own facts.” These words of <a href="https://www.vanityfair.com/news/2010/11/moynihan-letters-201011">Daniel Patrick Moynihan</a>, advisor to four presidents, could be the mantra for those trying to keep science from being drowned by misinformation.</p>
<p>[<em>You’re smart and curious about the world. So are The Conversation’s authors and editors.</em> <a href="https://theconversation.com/us/newsletters?utm_source=TCUS&utm_medium=inline-link&utm_campaign=newsletter-text&utm_content=youresmart">You can read us daily by subscribing to our newsletter</a>.]</p><img src="https://counter.theconversation.com/content/132158/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Chris Impey does not work for, consult, own shares in or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>The internet has allowed pseudoscience to flourish. Artificial intelligence could help steer people away from the bad information.Chris Impey, University Distinguished Professor of Astronomy, University of ArizonaLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1318322020-02-27T18:54:27Z2020-02-27T18:54:27ZFriday essay: a real life experiment illuminates the future of books and reading<p>Books are always transforming. The book we hold today has arrived through a number of materials (clay, papyrus, parchment, paper, pixels) and forms (tablet, scroll, codex, kindle). </p>
<p>The book can be a tool for communication, reading, entertainment, or learning; an object and a status symbol. </p>
<p>The most recent shift, from print media to digital technology, began around the middle of the 20th century. It culminated in two of the most ambitious projects in the history of the book (at least if we believe the corporate hype): the <a href="https://www.researchgate.net/publication/27710717_The_Google_Mass_Digitisation_Project_at_Oxford">mass-digitisation of books by Google</a> and the <a href="https://www.theatlantic.com/technology/archive/2019/08/amazons-plan-take-over-world-publishing/595630/">mass-distribution of electronic books by Amazon</a>.</p>
<p>The survival of bookshops and flourishing of libraries (in real life) defies predictions that the “<a href="https://www.theatlantic.com/magazine/archive/1994/09/the-end-of-the-book/376361/">end of the book</a>” is near. But even the most militant bibliophile will acknowledge how digital technology has called the “idea” of the book into question, once again.</p>
<p>To explore the potential for human-machine collaboration in reading and writing, we built a machine that makes poetry from the pages of any printed book. Ultimately, this project attempts to imagine the future of the book itself.</p>
<h2>A machine to read books</h2>
<p>Our custom-coded reading-machine reads and interprets real book pages, to create a new “<a href="https://www.ancient.eu/Illuminated_Manuscripts/">illuminated</a>” book of poetry.</p>
<p>The reading-machine uses Computer Vision and Optical Character Recognition to identify the text on any open book placed under its dual cameras. It then uses Machine Learning and Natural Language Processing technology to “read” the text for meaning, in order to select a short poetic combination of words on the page which it saves by digitally erasing all other words on the page.</p>
<p>Armed with this generated verse, the reading-machine searches the internet for an image – often a doodle or meme, which someone has shared and which has been stored in Google Images – to illustrate the poem.</p>
<p>Once every page in the book has been read, interpreted, and illustrated, the system publishes the results using an online printing service. The resulting volume is then added to a growing archive we call <a href="https://computervisionart.com/pieces2019/the-library-of-nonhuman-books/">The Library of Nonhuman Books</a>. </p>
<p>From the moment our machine completes its reading until the delivery of the book, our automated-art-system proceeds algorithmically – from interpreting and illuminating the poems, to pagination, cover design and finally adding the endmatter. This is all done without human intervention. The algorithm can generate a seemingly infinite number of readings of any book.</p>
<h2>The poetry</h2>
<p>The following poems were produced by the reading-machine from popular texts:</p>
<blockquote>
<p>deep down men try there </p>
<p>he’s large naked she’s even </p>
<p>while facing anything. </p>
</blockquote>
<p>from E.L. James’ <a href="https://www.goodreads.com/book/show/10818853-fifty-shades-of-grey?from_search=true&qid=SOwkpx3ir3&rank=1">Fifty Shades of Grey</a></p>
<blockquote>
<p>how parties popcorn </p>
<p>jukebox bathrooms depressed </p>
<p>shrug, yeah? all.</p>
</blockquote>
<p>from Bret Easton Ellis’ <a href="https://www.goodreads.com/book/show/9912.The_Rules_of_Attraction?from_search=true&qid=ZmniLWwydW&rank=1">The Rules of Attraction</a></p>
<blockquote>
<p>Oh and her bedroom </p>
<p>bathroom brushing sending it </p>
<p>garter too face hell.</p>
</blockquote>
<p>from Truman Capote’s <a href="https://www.goodreads.com/book/show/251688.Breakfast_at_Tiffany_s">Breakfast at Tiffany’s</a></p>
<h2>My algorithm, my muse</h2>
<p>So what does all this have to do with the mass-digitisation of books? </p>
<p>Faced with growing resistance from authors and publishers concerned with Google’s management of copyright, the infoglomerate pivoted away from its primary goal of providing a free corpus of books (a kind of modern day <a href="https://www.ancient.eu/article/207/what-happened-to-the-great-library-at-alexandria/">Library of Alexandria</a>) and towards a more modest index system used for searching inside the books Google had scanned. Google would now serve only short “snippets” of words highlighted on the original page. </p>
<p>Behind the scenes, Google had identified a different use for the texts. Millions of scanned books could be used in a field called <a href="https://becominghuman.ai/a-simple-introduction-to-natural-language-processing-ea66a1747b32">Natural Language Processing</a>. NLP allows computers to communicate with people using everyday language rather than code. The books originally scanned for humans were made available to machines for learning, and later imitating, human language.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/316593/original/file-20200221-92518-pegqqs.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Imagine infinite readings of the books we already have.</span>
<span class="attribution"><a class="source" href="https://unsplash.com/photos/Y7d265_7i08">Unsplash</a>, <a class="license" href="http://creativecommons.org/licenses/by/4.0/">CC BY</a></span>
</figcaption>
</figure>
<p>Algorithmic processes like NLP and <a href="https://emerj.com/ai-glossary-terms/what-is-machine-learning/">Machine Learning</a> hold the promise (or threat) of deferring much of our everyday reading to machines. History has shown that once machines know how to do something, we generally <a href="https://www.sciencedirect.com/science/article/pii/0166497294901015">leave them to it</a>. The extent to which we do this will depend on how much we value reading. </p>
<p>If we continue to defer our reading (and writing) to machines, we might make literature with our artificially intelligent counterparts. What will poetry become, with an algorithm as our muse? </p>
<p>We already have clues to this: from the almost obligatory use of emojis or Japanese <a href="http://kaomoji.ru/en/">Kaomoji</a> (顔文字) as visual shorthand for the emotional intent of our digital communication, to the layered meanings of internet memes, to the auto-generation of “<a href="https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction">fake news</a>” stories. These are the image-word hybrids we find in post-literate social media.</p>
<h2>To hide a leaf</h2>
<blockquote>
<p>Take the book, my friend, and read your eyes out, you will never find there what I find. </p>
</blockquote>
<p>Ralph Waldo Emerson’s <a href="https://archive.vcu.edu/english/engweb/transcendentalism/authors/emerson/essays/spirituallaws.html">Spiritual Laws</a> </p>
<p>Emerson’s challenge highlights the subjectivity we bring to reading. When we started working on the reading-machine we focused on discovering patterns of words within larger bodies of texts that have always been there, but have remained “hidden in plain sight”. Every attempt by the reading-machine generated new poems, all of them made from words that remained in their original positions on the pages of books.</p>
<p>The notion of a single book consisting of infinite readings is not new. We originally conceived our reading-machine as a way of making a mythical <a href="https://www.newyorker.com/magazine/1976/10/25/the-book-of-sand">Book of Sand</a>, described by Jorge Luis Borges in his 1975 parable.</p>
<p>Borges’ story is about the narrator’s encounter with an endless book which continuously recombines its words and images. Many have compared this impossible book to the internet of today. Our reading-machine, with the turn of each page of any physical book, calculates combinations of words on that page which, until that moment, have been seen, but not consciously perceived by the reader.</p>
<p>The title of our early version of the work was To Hide a Leaf. It was generated by chance when a prototype of the reading-machine was presented with a page from a book of Borges’ stories. The complete sentence from which the words were taken is: </p>
<blockquote>
<p>Somewhere I recalled reading that the best place to hide a leaf is in a forest. </p>
</blockquote>
<p>The latent verse our machine attempts to reveal in books also hides in plain sight, like a leaf in a forest; and the idea is also a play on a page being generally referred to as a “leaf of a book”. </p>
<p>Like the Book of Sand, perhaps all books can be seen as <a href="http://www.athrowofthedicewillneverabolishchance.com/">combinatorial machines</a>. We believed we could write an algorithm that could unlock new meanings in existing books, using only the text within that book as the key.</p>
<p>Philosopher Boris Groys described the result of the mass-digitisation of the book as <a href="https://www.amazon.com/Boris-Groys-Thoughts-Documenta-Gedanken/dp/3775728953">Words Without Grammar</a>, suggesting clouds of disconnected words.</p>
<p>Our reading-machine, and the Library of Nonhuman Books it is generating, is an attempt to imagine the book to come after these clouds of “words without grammar”. We have found the results are sometimes comical, often nonsensical, occasionally infuriating and, every now and then, even poetic.</p>
<figure>
<iframe src="https://player.vimeo.com/video/363445600" width="500" height="281" frameborder="0" webkitallowfullscreen="" mozallowfullscreen="" allowfullscreen=""></iframe>
<figcaption><span class="caption">Now that machines can read, will we defer the task to them?</span></figcaption>
</figure>
<hr>
<p><em>The reading-machine will be on display at the <a href="https://www.ngv.vic.gov.au/whats-on/programs-events/art-book-fair/">Melbourne Art Book Fair</a> in March and will collect a <a href="https://tokyotypedirectorsclub.org/en/news/tdc2020-results/">Tokyo Type Directors Club Award</a> in April. Nonhuman Books are available via <a href="http://www.atomicactivity.com/books/">Atomic Activity Books</a>.</em></p><img src="https://counter.theconversation.com/content/131832/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Andy Simionato is founder and editor of Atomic Activity Books, and is a lecturer at the School of Design, RMIT University.</span></em></p><p class="fine-print"><em><span>Karen ann Donnachie is founder and editor of Atomic Activity Books, an independent, experimental publishing concern. </span></em></p>We created a reading-machine that finds poetry hidden in plain sight in popular books. In doing so, we are exploring Natural Language Processing, Machine Learning and reading in a digitised world.Andy Simionato, Lecturer, RMIT UniversityKaren ann Donnachie, Independent artist / Lecturer (adjunct), RMIT UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1201992019-08-14T21:46:05Z2019-08-14T21:46:05ZThe language gives it away: How an algorithm can help us detect fake news<figure><img src="https://images.theconversation.com/files/284418/original/file-20190717-173334-1b9vdud.jpg?ixlib=rb-1.1.0&rect=0%2C0%2C6490%2C3957&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">In an attempt to address the growing problem of fake news online, an algorithm that identifies patterns in language may help distinguish between factual and inaccurate news articles.</span> <span class="attribution"><span class="source">Shutterstock</span></span></figcaption></figure><p>Have you ever read something online and shared it among your networks, only to find out it was false? </p>
<p>As a software engineer and computational linguist who spends most of her work and even leisure hours in front of a computer screen, I am concerned about what I read online. In the age of social media, <a href="http://www.digitalnewsreport.org/">many of us consume unreliable news sources</a>. We’re exposed to a wild flow of information in our social networks — especially if we spend a lot of time scanning our friends’ random posts on Twitter and Facebook. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/how-fake-news-gets-into-our-minds-and-what-you-can-do-to-resist-it-114921">How fake news gets into our minds, and what you can do to resist it</a>
</strong>
</em>
</p>
<hr>
<p>My colleagues and I at the <a href="http://www.sfu.ca/discourse-lab.html">Discourse Processing Lab at Simon Fraser University</a> have conducted research on the linguistic characteristics of fake news.</p>
<h2>The effects of fake news</h2>
<p>A study in the United Kingdom found that <a href="https://doi.org/10.1145/3173574.3173950">about two-thirds of the adults surveyed regularly read news on Facebook, and that half of those had the experience of initially believing a fake news story</a>. Another study, conducted by researchers at the Massachusetts Institute of Technology, focused on the cognitive aspects of exposure to fake news and <a href="https://doi.org/10.1016/j.cognition.2018.06.011">found that, on average, newsreaders believe a false news headline at least 20 percent of the time</a>. </p>
<p>False stories are now <a href="http://doi.org/10.1126/science.aap9559">spreading 10 times faster than real news</a> and the problem of fake news seriously <a href="https://theconversation.com/the-real-consequences-of-fake-news-81179">threatens our society</a>. </p>
<p>For example, during the 2016 election in the United States, an astounding number of U.S. citizens believed and shared a patently false conspiracy claiming that Hilary Clinton was connected to <a href="https://www.nytimes.com/interactive/2016/12/10/business/media/pizzagate.html">a human trafficking ring run out of a pizza restaurant</a>. The owner of the restaurant received death threats, and one believer showed up in the restaurant with a gun. This — and <a href="https://www.nature.com/articles/s41467-018-07761-2">a number of other fake news stories</a> distributed during the election season — had an undeniable impact on people’s votes.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/trump-may-owe-his-2016-victory-to-fake-news-new-study-suggests-91538">Trump may owe his 2016 victory to 'fake news,' new study suggests</a>
</strong>
</em>
</p>
<hr>
<p>It’s often difficult to find the origin of a story after partisan groups, social media bots and friends of friends <a href="https://www.cits.ucsb.edu/fake-news/spread">have shared it thousands of times</a>. Fact-checking websites such as <a href="https://www.snopes.com/fact-check/">Snopes</a> and <a href="https://www.buzzfeed.com/badge/fact-checker">Buzzfeed</a> can only address a small portion of the most popular rumors. </p>
<p>The technology behind the internet and social media has enabled this spread of misinformation; maybe it’s time to ask what this technology has to offer in addressing the problem.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/yc8EZ3kZ1zQ?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">In an interview, Hilary Clinton discusses ‘Pizzagate’ and the problem of fake news online.</span></figcaption>
</figure>
<h2>Giveaways in writing style</h2>
<p>Recent advances in machine learning have made it possible for computers to instantaneously complete tasks that would have taken humans much longer. For example, there are computer programs that help police identify criminal faces <a href="https://www.nbcnews.com/news/us-news/how-facial-recognition-became-routine-policing-tool-america-n1004251">in a matter of seconds</a>. This kind of artificial intelligence trains algorithms to classify, detect and make decisions. </p>
<p>When machine learning is applied to natural language processing, it is possible to build text classification systems that recognize one type of text from another.</p>
<p>During the past few years, natural language processing scientists have become more active in building algorithms to detect misinformation; this helps us to understand the characteristics of fake news and develop technology to help readers. </p>
<p>One approach finds relevant sources of information, assigns each source a credibility score and then integrates them to confirm or debunk a given claim. This approach is heavily dependent on <a href="http://news.mit.edu/2018/mit-csail-machine-learning-system-detects-fake-news-from-source-1004">tracking down the original source of news and scoring its credibility based on a variety of factors</a>.</p>
<p>A second approach examines the writing style of a news article rather than its origin. The linguistic characteristics of a written piece can tell us a lot about the authors and their motives. For example, specific words and phrases tend to <a href="https://arxiv.org/abs/1708.07104">occur more frequently in a deceptive text compared to one written honestly</a>.</p>
<h2>Spotting fake news</h2>
<p>Our research identifies linguistic characteristics to detect fake news using machine learning and natural language processing technology. Our analysis of a large collection of fact-checked news articles on a variety of topics shows that, on average, fake news articles use more expressions that are common in hate speech, as well as words related to sex, death and anxiety. Genuine news, on the other hand, contains a larger proportion of words related to work (business) and money (economy).</p>
<p>This suggests that a stylistic approach combined with machine learning might be useful in detecting suspicious news. </p>
<p>Our fake news detector is built based on linguistic characteristics extracted from a large body of news articles. It takes a piece of text and shows how similar it is to the fake news and real news items that it has seen before. (<a href="http://fake-news.research.sfu.ca">Try it out!</a>)</p>
<p>The main challenge, however, is to build a system that can handle the vast variety of news topics and the quick change of headlines online, because computer algorithms learn from samples and if these samples are not sufficiently representative of online news, <a href="http://doi.org/10.1177/2053951719843310">the model’s predictions would not be reliable</a>.</p>
<p>One option is to have human experts collect and label a large quantity of fake and real news articles. This data enables a machine-learning algorithm to find common features that keep occurring in each collection regardless of other varieties. Ultimately, the algorithm will be able to distinguish with confidence between previously unseen real or fake news articles.</p>
<p>[ <em>Deep knowledge, daily.</em> <a href="https://theconversation.com/ca/newsletters?utm_source=TCCA&utm_medium=inline-link&utm_campaign=newsletter-text&utm_content=deepknowledge">Sign up for The Conversation’s newsletter</a>. ]</p><img src="https://counter.theconversation.com/content/120199/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Fatemeh Torabi Asr does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>Using machine learning and natural language processing, researchers are developing an algorithm that can distinguish between real and fake news articles.Fatemeh Torabi Asr, Postdoctoral research fellow, Simon Fraser UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/925652018-03-19T12:15:08Z2018-03-19T12:15:08ZIn the #MeToo era, a chatbot can help people report workplace harassment<figure><img src="https://images.theconversation.com/files/210087/original/file-20180313-30972-1emiu0v.jpg?ixlib=rb-1.1.0&rect=599%2C314%2C3410%2C2362&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><a class="source" href="https://www.shutterstock.com/download/confirm/1009471495?size=huge_jpg">Shutterstock</a></span></figcaption></figure><p>Campaigns like <a href="https://theconversation.com/sexual-violence-may-be-in-the-hollywood-spotlight-but-there-are-limits-to-speaking-out-92090">#MeToo and Time’s Up</a> mean that public discussion about <a href="https://theconversation.com/how-to-recognise-and-start-tackling-sexual-harassment-in-the-workplace-86235">sexual harassment</a> has finally bubbled up to the surface. The movements also highlight how such disturbing incidents have routinely gone unreported or been outright ignored.</p>
<p>To find out the extent of the problem, our team at the universities of Sussex, Derby and UCL recently conducted a review on the lack of reporting of workplace harassment and discrimination, by summarising the past 18 years of relevant research in this area.</p>
<p>The findings informed our development of a <a href="https://theconversation.com/uk/search?utf8=%E2%9C%93&q=chatbot">chatbot</a>, which uses a computer-controlled robot to mimic a human interviewer, for recording and reporting workplace harassment anonymously. <a href="https://talktospot.com/">Spot</a> is powered by <a href="https://theconversation.com/uk/search?utf8=%E2%9C%93&q=natural+language+processing">natural language processing</a> – a basic branch of <a href="https://theconversation.com/uk/topics/artificial-intelligence-90">artificial intelligence</a> and <a href="https://theconversation.com/uk/search?utf8=%E2%9C%93&q=computer+science">computer science</a> – which understands and interprets human language.</p>
<p>The Spot chatbot – designed to improve the reporting process for those who have experienced or witnessed workplace harassment or discrimination – guides the participant through an evidence-based <a href="https://crestresearch.ac.uk/comment/the-cognitive-interview/">cognitive interview</a> without requiring them to talk to a human. This is used to get a high quality account of what happened that avoids leading questions that can produce false memories.</p>
<p>It also removes barriers to harassment reporting, because participants can log incidents without talking to a human. This bypasses concerns about trust, confidentiality and doubts being cast over harassment allegations.</p>
<p>Spot gives the participant a time-stamped report at the end of the session. Users can keep this as evidence in case it is needed later. It can also send an anonymous version of the report to the user’s employer at the request of the participant. The report is <a href="https://talktospot.com/faq#safety">deleted from Spot’s servers</a> 30 days after it is downloaded by the participant.</p>
<p>The growing awareness of the sexism and power dynamics that underpin toxic workplaces shows why it’s time to start a broader dialogue about harassment and discrimination, and to find better ways of tackling this issue.</p>
<p>With our research colleague, <a href="https://www.talktospot.com/about-us">Rashid Minhas</a>, we broadly reviewed preliminary research into all forms of workplace discrimination to try to understand how much harassment is experienced, how much is reported and what the repercussions of reporting were for people speaking out against abuse. Different studies approached the same issue in fundamentally different ways, making them difficult to generalise. Nonetheless, we found some startling figures.</p>
<h2>Key findings</h2>
<p>We identified six main problems associated with reporting workplace harassment and discrimination and evidenced by studies:</p>
<ol>
<li><p>Workplace harassment is rarely reported. The most comprehensive estimate from a sample of 91,503 filed charges in the US found that about <a href="https://www.eeoc.gov/eeoc/task_force/harassment/upload/rebooting_harassment_prevention.pdf">70%</a> of harassment cases went unreported. However, most studies found reporting rates that were far lower than this. In some industries, such as the <a href="http://www.msnbc.com/msnbc/report-sexual-harassment-rampant-the-restaurant-industry">restaurant business</a>, just 3% of harassment and discrimination is logged.</p></li>
<li><p>There are many barriers to reporting. When asked why, people stated that they <a href="http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0102172">did not know how to report</a> harassment allegations; they feared retaliation citing <a href="http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0102172">negative consequences</a> when others reported incidents; they were <a href="https://www.researchgate.net/profile/Belle_Ragins/publication/6200999_Making_the_Invisible_Visible_Fear_and_Disclosure_of_Sexual_Orientation_at_Work/links/56b3eeea08ae5deb2657e71a.pdf">reluctant to disclose</a> personal issues or protected characteristics (like their sexuality); and they had concerns about <a href="http://www.tandfonline.com/doi/full/10.1080/23311908.2015.1121066">trust</a> and confidentiality.</p></li>
<li><p>Harassment is linked to poor health. Studies show that the <a href="https://www.jstage.jst.go.jp/article/joh/51/4/51_L8143/_pdf">health</a> of harassed people declines, and how harassment is related to an increase in <a href="http://oem.bmj.com/content/57/10/656.long">sickness absence from work</a>.</p></li>
<li><p>No <a href="http://doi.apa.org/getdoi.cfm?doi=10.1037/1076-8998.8.4.247">health-related consequences</a> have been found when reports were handled appropriately and management responses were prompt. In other words, it was not always the harassment itself that was related to poor health, but the ways in which complaints were <a href="http://doi.apa.org/getdoi.cfm?doi=10.1037/0021-9010.87.2.230">handled</a>.</p></li>
<li><p>Harassment is linked to negative workplace outcomes, studies show. Harassed people were <a href="http://doi.apa.org/getdoi.cfm?doi=10.1037/0021-9010.90.3.483">less satisfied</a> with their jobs, and were more likely to want to <a href="https://lra.le.ac.uk/bitstream/2381/4362/1/pserc01-2.pdf">leave</a>. Harassment was also found to have led to bad press or <a href="https://www.eeoc.gov/eeoc/plan/upload/2016par.pdf">litigation</a>.</p></li>
<li><p>A positive organisational structure helped prevent harassment and facilitate reporting. Negative issues could be addressed by improving <a href="http://doi.apa.org/getdoi.cfm?doi=10.1037/0021-9010.86.6.1244">workplace culture</a>, including showing <a href="http://www.journals.cambridge.org/abstract_S0144686X12000438">respect and support</a>.</p></li>
</ol>
<h2>Speak out</h2>
<p>Chatbots can explain what needs to happen if a person has been harassed or discriminated against. They can tackle otherwise difficult conversations. They don’t judge, and they don’t have the same biases that humans do. They are available online day and night. They can allow for confidentiality and anonymity, reducing potential backlash. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=439&fit=crop&dpr=1 600w, https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=439&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=439&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=551&fit=crop&dpr=1 754w, https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=551&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/210090/original/file-20180313-30954-kg9d4u.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=551&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Spot cofounders Dylan Marriott, Daniel Nicolae and Julia Shaw.</span>
<span class="attribution"><a class="source" href="https://talktospot.com/about-us">Spot</a></span>
</figcaption>
</figure>
<p>The bots can be designed to ask the best questions every time, although not all do. <a href="https://theconversation.com/microsofts-racist-chatbot-tay-highlights-how-far-ai-is-from-being-truly-intelligent-56881">Microsoft’s chatbot Tay</a> – a machine learning experiment in which it was hoped that the bot would appear as a young woman interacting with others on Twitter – serves as a cautionary tale. It spectacularly backfired when Tay started tweeting Nazi, anti-feminist and racist views, after the project was hijacked by miscreants.</p>
<p>However, it has become clear from our research, and the #MeToo campaign, that many people want to speak up about harassment at work but are still reluctant to report it. Chatbots can offer reliable and scalable services to empower people to come forward by removing some of the barriers to reporting.</p>
<p>More voices must join the conversation to speak out against harassment – only together can we change the institutional structures that have allowed it to flourish. We hope that Spot, and tools like it, will help facilitate improved reporting in the workplace.</p><img src="https://counter.theconversation.com/content/92565/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Julia Shaw is one of the cofounders of Spot.</span></em></p><p class="fine-print"><em><span>Camilla Elphick is a researcher at Spot.</span></em></p>Spot removes traditional barriers to reporting abusive behavior, because participants can log incidents without talking to a human.Julia Shaw, Research associate, UCLCamilla Elphick, Psychology PhD candidate,, University of SussexLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/829362017-09-05T10:59:24Z2017-09-05T10:59:24ZWe need to consider the social implications of bots writing books about instant chocolate milk<figure><img src="https://images.theconversation.com/files/184566/original/file-20170904-9750-1qtwf2j.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><span class="source">kikujungboy / Shutterstock.com</span></span></figcaption></figure><p>We still tend to think of authors as geniuses. Conversations about prize-winning novels often revolve around authorial intent, and writing a bestseller will tend to confer a certain amount of fame. But this timeworn state of affairs is shifting. The reader, more than ever, is moving centre stage – a shift that is being catalysed by the development of natural language generation (NLG) systems by computer scientists around the world. Written language is no longer a uniquely human construct.</p>
<p>NLG is the process whereby computer data is translated into everyday human languages. NLG has already been applied for many purposes, including aesthetics <a href="https://www.theguardian.com/technology/2016/may/17/googles-ai-write-poetry-stark-dramatic-vogons">(poetry, say)</a> and information transfer <a href="http://ieeexplore.ieee.org/document/294135">(for example, weather updates)</a>. We’re now at a point where computers generate texts largely indistinguishable from human-authored texts, and at rates incomparable to that of humans.</p>
<p>Written language has, for hundreds of years, served as much of the world’s dominant form of knowledge transfer. Yet history reveals that the author hasn’t always been regarded as an individual creative genius. The medieval European writer, for example, was what we’d now deem a plagiarist, pulling from a range of source texts to make their own. A text wasn’t a manifestation of individuality, but a means for preserving knowledge passed through the generations. Medieval scholars cared little about their books’ writers: they cared about the ancient truths held within.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=799&fit=crop&dpr=1 600w, https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=799&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=799&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=1004&fit=crop&dpr=1 754w, https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=1004&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/184688/original/file-20170905-13755-lcio5k.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=1004&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">1628 depiction of Haarlem printing press from 1440.</span>
<span class="attribution"><a class="source" href="https://commons.wikimedia.org/w/index.php?search=printing+press+1440&title=Special:Search&profile=images&fulltext=1&searchToken=uyy00391r9fn4ucn3336710z">Wikimedia Commons</a></span>
</figcaption>
</figure>
<p>However, ushering in what book historian <a href="https://books.google.co.uk/books/about/The_Printing_Press_as_an_Agent_of_Change.html?id=0-FThHK2DNMC&redir_esc=y">Elisabeth Eisenstein calls</a> “cults of personality”, print supported a shift in the cultural mindset. Suddenly the author’s individuality, more than their contribution to the collective, was praised. Of course, print wasn’t solely responsible for this large-scale change, but it certainly both influenced and was influenced by the development of modern individuality. </p>
<p>Now, the success of modern NLG companies represents a further shift in the same direction, further emphasising the individual over the collective. The reader dominates when texts are created especially for her.</p>
<h2>The cult of the individual</h2>
<p>NLG supports the production of highly localised content: so localised that the algorithm produces it for the individual rather than any collective, accommodating unique social circumstances and worldviews. In such a hyper-individualised climate, reading becomes less of an effort to find authorial intention, and more a personal experience of meaning-making. This shift is symptomatic of a greater shift from a shared culture that reveres expert views afforded by social institutions to one comprising niche groups representing personal interests. No one likes being told what to think. NLG can accommodate individuality.</p>
<p>Nowhere is NLG’s affirmation of individualisation clearer than in the rise of companies like <a href="http://automatedinsights.com">Automated Insights</a> and <a href="http://narrativescience.com">Narrative Science</a>, which specialise in the generation of news and business intelligence reports for large-scale and niche audiences alike.</p>
<p>Transforming datasets (for example, sports scores, performance measures) into readable narratives, these companies’ systems rapidly generate texts that are highly personalised in both content and register. Such texts present situations in which information appears disembodied from its conveyor. The author is obscured, an uncertain entanglement of human and computer.</p>
<p>While NLG systems may produce texts for mass audiences, one of these systems’ most novel features is their ability to generate texts on even the most niche subjects. Philip Parker’s patented algorithms, for example, generate entire books marketed to specialised clientele. The 2018-2023 World Outlook for Instant Chocolate Milk, Weight Control Products, Whole Milk Powder, Malted Milk Powder, and Other Dry Milk Products Shipped in Consumer Packages Weighing Three Pounds of Less Excluding Nonfat Dry Milk and <a href="https://www.amazon.com/2018-2023-Outlook-Instant-Chocolate-Products/dp/B06XD91YDC">Infants’ Formula</a> is 310 pages long and priced at a reasonable $995 USD. This book – like the rest of the series – offers easy-to-read charts revealing industry trends. Interested in instant chocolate milk? Parker’s book is a godsend.</p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=420&fit=crop&dpr=1 600w, https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=420&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=420&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=528&fit=crop&dpr=1 754w, https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=528&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/184695/original/file-20170905-13747-1b7d0qo.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=528&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">A bot can write a book about that.</span>
<span class="attribution"><span class="source">urbanbuzz / Shutterstock.com</span></span>
</figcaption>
</figure>
<h2>Intellectual blinders</h2>
<p>But there are ethical issues with such “personalisation algorithms” driving niche content. After all, there’s a fine line between supporting readers’ decisions and controlling them, as a group of researchers from the Oxford Internet and Alan Turing Institutes <a href="http://journals.sagepub.com/doi/abs/10.1177/2053951716679679">argue</a>. In theory, personalisation improves decision-making by filtering out irrelevant information
contributing to information overload. </p>
<p>But this filtering process, which is always subjective, reduces exposure to diverse views that may be considered irrelevant or contradictory to the individual being catered for. And if everyone is presented with customised content, exaggerated monocultures form wherein individuals become trapped in echo chambers that only reinforce their current beliefs instead of challenging them.</p>
<p>It may seem pedantic to dedicate such consideration to personalised content like business reports. Yet the systematic omission of “irrelevant” data creates intellectual blinders with very real social implications. We need balanced perspectives – the products of diverse views that challenge our own – to create policies that serve wide publics. </p>
<p>So today’s tendency towards populist anti-intellectualism not only stems from underlying distrust of institutionalised hierarchy, but also from the rise of a hyper-individualised culture partially supported by personalisation algorithms. Everyone’s opinion is valid in light of the information to which she’s been exposed. Indeed, individualised computer-generated content seems to indicate a general defiance of institutionalised hierarchy. The reader, rather than the author, is in control. The reader determines what information matters.</p>
<p>There’s value in individualised computer-generated texts. Readers can get the information they want, quickly. Businesses can evaluate performance in seconds. The issue is not the technology – it is what the technology represents. NLG companies promoting mass production of customised texts survive because readers believe that their individualities are, at least in some contexts, to be prioritised over the collective. This is nothing new: the emergence of print in the 15th century catalysed the development of the modern individual. What is new is the state of hyper-individualisation permitted by an increasingly digitised – and customised – culture. </p>
<p>Individuality is important. But so is society. There is a balance to be struck as the use of NLG continues to spread.</p><img src="https://counter.theconversation.com/content/82936/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Leah Henrickson does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>The mass generation of customised content may have very real social implications.Leah Henrickson, Doctoral Student, Loughborough UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/714762017-01-19T19:08:56Z2017-01-19T19:08:56ZAustralians on Twitter were ‘positive’ about the US election of Donald Trump<p>Australians reacted more “positive” than “negative” to the election of Donald Trump as the next president of the United States, according to a sentiment analysis study of tweets that were posted at the time.</p>
<p>Only tweets sent on November 10, 2016, (just after the result of the US election) that included the word “Trump” and were sent from an Australian capital city were analysed.</p>
<p>This resulted in 32,908 tweets including retweets being retrieved. For the purpose of this analysis we classified the tweet sentiment as either positive, negative or neutral.</p>
<iframe src="https://datawrapper.dwcdn.net/qXIVb/2/" frameborder="0" allowtransparency="true" allowfullscreen="allowfullscreen" webkitallowfullscreen="webkitallowfullscreen" mozallowfullscreen="mozallowfullscreen" oallowfullscreen="oallowfullscreen" msallowfullscreen="msallowfullscreen" width="100%" height="304"></iframe>
<p>The figures (above) display the sentiment for each capital city and show that in Sydney, Brisbane, Canberra and Hobart there were more positive tweets about Trump. In Darwin, Adelaide, Melbourne and Perth there were more negative tweets.</p>
<p>But counted overall, 48.63% of the tweets were considered positive compared to 44.65% negative and 6.72% neutral.</p>
<h2>More detail in the tweets</h2>
<p>To try to get a better understanding of the divided sentiment, individual tweets were investigated and it soon became apparent that the sentiment analysis had difficulty identifying sarcasm and humour.</p>
<p>For example, tweets that included “LOL” (Laughing Out Loud) were interpreted as a strong positive sentiment where in many cases it was not.</p>
<p><div data-react-class="Tweet" data-react-props="{"tweetId":"798040815700320256"}"></div></p>
<p>In analysis of another tweet “Dear Harvard Business School: don’t normalize Trump’s rise to power. Fascism is not a "marketing strategy”. <a href="https://twitter.com/HarvardHBS">@HarvardHBS</a> <a href="https://twitter.com/HBSWK">@HBSWK</a>“, from Melbourne’s <a href="https://twitter.com/creatrixtiara">@creatrixtiara</a> (above), it was determined that this was a request rather than a sentiment.</p>
<p><div data-react-class="Tweet" data-react-props="{"tweetId":"798364639117459456"}"></div></p>
<p>The system used could not determine the sentiment of a tweet by Brisbane’s <a href="https://twitter.com/JamesPinnell">@JamesPinnell</a>: "The only thing that gives me a tiny inkling of hope is that Trump’s kids seem to actually be fairly bright and also in his ear a lot.”</p>
<h2>It’s the popular words that count</h2>
<p>We performed an analysis of the most popular words and terms used in the positive and negative sentiment tweets, in order to get a better insight into the intended sentiments.</p>
<p>The analysis shows the disparity of views and sentiments which have characterised this election.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=301&fit=crop&dpr=1 600w, https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=301&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=301&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=378&fit=crop&dpr=1 754w, https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=378&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/153363/original/image-20170119-26550-enyxc6.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=378&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Positive sentiment: most frequent words tweeted.</span>
<span class="attribution"><span class="license">Author provided</span></span>
</figcaption>
</figure>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=322&fit=crop&dpr=1 600w, https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=322&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=322&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=404&fit=crop&dpr=1 754w, https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=404&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/153364/original/image-20170119-26550-1dlvedg.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=404&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Negative sentiment: most frequent words tweeted.</span>
<span class="attribution"><span class="license">Author provided</span></span>
</figcaption>
</figure>
<h2>The tools of the study</h2>
<p>This study was made possible due to recent advances in business analytic tools, with Victoria University’s Business Analytics and Big Data Lab working in partnership with SAP. </p>
<p>Traditionally business analytic tools focus on structured data to gain insight and facilitate decision making. This type of data is contained in databases and spreadsheets and is characterised by a combination of fields in a record. Structured data has the advantage of being easily entered, stored, queried and analysed.</p>
<p>But much of the data that is contained in social media, including tweets, is referred to as unstructured data. It doesn’t reside in fields and record structures and so it’s difficult to analyse using traditional methods. </p>
<p>A technique referred to as Text Analysis is the process of analysing unstructured text to extract relevant information and then transform that information into a structured format for analysis. </p>
<p>Text Analysis uses Natural Language Processing to linguistically understand the text and apply statistical techniques to facilitate the analyses. </p>
<p>SAP’s <a href="https://www.sap.com/product/technology-platform/hana.html">HANA</a> database platform can search, analyse and mine text. It allowed us to perform a traditional exact string search such as “Trump is wonderful” or a fuzzy search (Google like) where text can be found irrespective of the sequence of words.</p>
<p>It also allowed us to provide meaning to the text through tokenisation and stemming.</p>
<p>For example in the text “Trump wins Florida in 2016”, SAP HANA would identify the entities of Trump as a Person, Florida as a State and 2016 as a Year.</p>
<p>This form of analysis can be further enhanced through the use of fact extraction. This is where rules are used as a basis to determine relationships between the identified entities.</p>
<h2>Finding the positives and negatives</h2>
<p>The most common form of fact extraction is sentiment analysis. A statement like “I love Trump” would be identified as a strong positive sentiment in relation to Trump. </p>
<p>The polarity of a statement can be identified (either strong or weak) in addition to the sentiment (either positive, negative or neutral).</p>
<p>A number of pre-defined rules exist in SAP HANA to facilitate sentiment analysis but these can be further extended through customisation of keyword dictionaries depending on the scenario.</p>
<p>For example the word “Trump” can be restricted to refer to only a person rather than an action (for instance playing a trump card). The sentiment analysis can be applied in ten different languages and can also identify requests, emoticons and profanity.</p>
<h2>Future studies</h2>
<p>It is obvious that any analysis needs to be treated with caution in regards to sarcasm, humour and other possible variations. As the tools improve, these issues will hopefully be addressed.</p>
<p>But the research provided an example of how natural language processing can be applied to social media to gain insight to the sentiment of a specific population in regards to an event, as well as to potential limitations.</p>
<p>We were also mindful that we were looking at tweets in Australia of an event that was happening elsewhere, in the US. We look forward to analysing the next election, possibly in Australia.</p>
<p>It would also interesting to see how people react once Trump is installed as the 45th US president. He has promised to continue using his personal Twitter handle <a href="https://twitter.com/realDonaldTrump">@realDonaldTrump</a> instead of <a href="https://twitter.com/POTUS">@POTUS</a>, used by outgoing 44th president, Barack Obama.</p><img src="https://counter.theconversation.com/content/71476/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Paul Hawking is affiliated with SAP Australian User Group </span></em></p><p class="fine-print"><em><span>Scott Bingley does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>On the eve of the inauguration of Donald Trump as the next US president, we take a look back at the Australian reaction on Twitter to his election.Paul Hawking, Associate Professor Information Systems, Victoria UniversityScott Bingley, Lecturer of Information Systems, Victoria UniversityLicensed as Creative Commons – attribution, no derivatives.