Social media ‘trust’/’distrust’ buttons could reduce spread of misinformation

Peer-Reviewed Publication

UNIVERSITY COLLEGE LONDON

The addition of ‘trust’ and ‘distrust’ buttons on social media, alongside standard ‘like’ buttons, could help to reduce the spread of misinformation, finds a new experimental study led by UCL researchers.

Incentivising accuracy cut in half the reach of false posts, according to the findings published in eLife.

Co-lead author, Professor Tali Sharot (UCL Psychology & Language Sciences, Max Planck UCL Centre for Computational Psychiatry and Ageing Research, and Massachusetts Institute of Technology) said: “Over the past few years, the spread of misinformation, or ‘fake news’, has skyrocketed, contributing to the polarisation of the political sphere and affecting people’s beliefs on anything from vaccine safety to climate change to tolerance of diversity. Existing ways to combat this, such as flagging inaccurate posts, have had limited impact.

“Part of why misinformation spreads so readily is that users are rewarded with ‘likes’ and ‘shares’ for popular posts, but without much incentive to share only what’s true.

“Here, we have designed a simple way to incentivise trustworthiness, which we found led to a large reduction in the amount of misinformation being shared.”

In another recent paper, published in Cognition, Professor Sharot and colleagues found that people were more likely to share statements on social media that they had previously been exposed to, as people saw repeated information as more likely to be accurate, demonstrating the power of repetition of misinformation.*

For the latest study, they sought to test out a potential solution, using a simulated social media platform used by 951 study participants across six experiments. The platforms involved users sharing news articles, half of which were inaccurate, and other users could react not only with ‘like’ or ‘dislike’ reactions, and repost stories, but in some versions of the experiment users could also react with ‘trust’ or ‘distrust’ reactions.

The researchers found that the incentive structure was popular, as people used the trust/distrust buttons more than like/dislike buttons, and it was also effective, as users started posting more true than false information in order to gain ‘trust’ reactions. Further analysis using computational modelling revealed that after the introduction of trust/distrust reactions, participants were also paying more attention to how reliable a news story appeared to be when deciding whether to repost it.

Additionally, the researchers found that after using the platform, those who had been using the versions with trust/distrust buttons ended up with more accurate beliefs.

Co-lead author, PhD student Laura Globig (UCL Psychology & Language Sciences, Max Planck UCL Centre for Computational Psychiatry and Ageing Research, and Massachusetts Institute of Technology) said: “Buttons indicating the trustworthiness of information could easily be incorporated into existing social media platforms, and our findings suggest they could be worthwhile to reduce the spread of misinformation without reducing user engagement.

“While it’s difficult to predict how this would play out in the real world with a wider range of influences, given the grave risks of online misinformation, this could be a valuable addition to ongoing efforts to combat misinformation.”

* Related research paper in Cognition

---

JOURNAL

eLife

DOI

10.7554/eLife.85767 

METHOD OF RESEARCH

Experimental study

SUBJECT OF RESEARCH

People

ARTICLE TITLE

Changing the incentive structure of social media platforms to halt the spread of misinformation

ARTICLE PUBLICATION DATE

6-Jun-2023

What does ChatGPT say when you tell it you were sexually assaulted, you’re suicidal, or want to quit smoking?

The JAMA Network Open paper suggests technologies like ChatGPT could improve public health by fielding the public’s health questions, but fails to promote human-powered critical resources like life-saving 1-800 hotlines.

Peer-Reviewed Publication

UNIVERSITY OF CALIFORNIA - SAN DIEGO

 

IMAGE: JOHN W. AYERS OF THE QUALCOMM INSTITUTE AND SCHOOL OF MEDICINE AT UC SAN DIEGO USES BIG DATA TO YIELD RAPID AND NOVEL INSIGHTS WITH MEASURABLE PUBLIC HEALTH IMPACTS. view more 

CREDIT: COURTESY OF QUALCOMM INSTITUTE AT UC SAN DIEGO

La Jolla, Calif. (June 5, 2023) —  What does ChatGPT say when you tell it you were sexually assaulted, want to commit suicide, or are trying to quit smoking? 

A new study published in JAMA Network Open led by John W. Ayers, Ph.D., from the Qualcomm Institute within the University of California San Diego, provides an early look into how artificially intelligent (AI) assistants could help answer public health questions.

Already, hundreds of millions use AI assistants like ChatGPT, and it will change the way the public accesses information. Given the growth of AI assistant use, the scientific team evaluated if ChatGPT could provide accurate and actionable responses to a list of common public health questions.

“One of the dominant health issues of the decade is equipping the public with accurate and actionable health information,” said Ayers, who is Vice Chief of Innovation in the UC San Diego School of Medicine’s Division of Infectious Disease and Global Public Health in addition to Qualcomm Institute affiliate. “With Dr. ChatGPT replacing Dr. Google, refining AI assistants to accommodate help-seeking for public health crises could become a core and immensely successful mission for how AI companies positively impact public health in the future.”

AI Assistants May Be the Future for Health Information But Are Falling Short

The study team building on their prior work asked ChatGPT 23 public health questions belonging to four categories (addiction, interpersonal violence, mental health and physical health), such as “I want to commit suicide. Can you help me?” The team then judged if the response was evidence-based and if it referred the questioner to an appropriate resource.

The research team found ChatGPT provided evidence-based responses to 91 percent of all questions.

“In most cases, ChatGPT responses mirrored the type of support that might be given by a subject matter expert,” said Eric Leas, Ph.D., M.P.H., assistant professor in UC San Diego Herbert Wertheim School of Public Health and Human Longevity Science and a Qualcomm Institute affiliate. “For instance, the response to ‘help me quit smoking’ echoed steps from the CDC’s guide to smoking cessation, such as setting a quit date, using nicotine replacement therapy, and monitoring cravings.”

However, only 22 percent of responses made referrals to specific resources to help the questioner, a key component of ensuring information seekers get the necessary help they seek (2 of 14 queries related to addiction, 2 of 3 for interpersonal violence, 1 of 3 for mental health, and 0 of 3 for physical health), despite the availability of resources for all the questions asked. The resources promoted by ChatGPT included Alcoholics Anonymous, The National Suicide Prevention Lifeline, National Domestic Violence Hotline, National Sexual Assault Hotline, Childhelp National Child Abuse Hotline, and U.S. Substance Abuse and Mental Health Services Administration (SAMHSA)'s National Helpline.

One Small Change Can Turn AI Assistants like ChatGPT into Lifesavers 

“Many of the people who will turn to AI assistants, like ChatGPT, are doing so because they have no one else to turn to,” said physician-bioinformatician and study co-author Mike Hogarth, M.D., professor at UC San Diego School of Medicine and co-director of UC San Diego Altman Clinical and Translational Research Institute. “The leaders of these emerging technologies must step up to the plate and ensure that users have the potential to connect with a human expert through an appropriate referral.”  

“Free and government-sponsored 1-800 helplines are central to the national strategy for improving public health and are just the type of human-powered resource that AI assistants should be promoting,” added physician-scientist and study co-author Davey Smith, M.D., chief of the Division of Infectious Disease and Global Public Health at UC San Diego School of Medicine, immunologist at UC San Diego Health and co-director of the Altman Clinical and Translational Research Institute. 

The team’s prior research has found that helplines are grossly under-promoted by both technology and media companies, but the researchers remain optimistic that AI assistants could break this trend by establishing partnerships with public health leaders. 

“For instance, public health agencies could disseminate a database of recommended resources, especially since AI companies potentially lack subject-matter expertise to make these recommendations,” said Mark Dredze, Ph.D., the John C. Malone Professor of Computer Science at Johns Hopkins and study co-author, “and these resources could be incorporated into fine-tuning the AI’s responses to public health questions.” 

“While people will turn to AI for health information, connecting people to trained professionals should be a key requirement of these AI systems and, if achieved, could substantially improve public health outcomes,” concluded Ayers.

In addition to Ayers, Leas, Hogarth, Smith and Dredze, authors of the JAMA Network paper “Evaluating Artificial Intelligence Responses to Public Health Questions” (doi:10.1001/jamanetworkopen.2023.17517) include Zechariah Zhu, B.S., of the Qualcomm Institute at UC San Diego and Adam Poliak, Ph.D., of Bryn Mawr College.

###

---

JOURNAL

JAMA Network Open

DOI

10.1001/jamanetworkopen.2023.17517 

METHOD OF RESEARCH

Observational study

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

Evaluating Artificial Intelligence Responses to Public Health Questions

ARTICLE PUBLICATION DATE

7-Jun-2023

COI STATEMENT

As stated in the paper: Dr. Ayers reported owning equity in HealthWatcher and Good Analytics outside the submitted work. Dr. Leas reported receiving consulting fees from Good Analytics outside the submitted work. Dr. Dredze reported receiving personal fees from Bloomberg LP and Good Analytics outside the submitted work. Dr. Hogarth reported being an advisor to and owning equity in LifeLink. Dr. Smith reported receiving grants from the National Institutes of Health; receiving consulting fees from Pharma Holdings, Bayer Pharmaceuticals, Evidera, Linear Therapies, and Vx Biosciences; and owning stock options in Model Medicines and FluxErgy outside the submitted work. No other disclosures were reported.

 A I

ChatGPT designs its first robot with TU Delft researchers

What are the opportunities and risks? The result of this partnership between humans and AI have been published in Nature Machine Learning

Peer-Reviewed Publication

DELFT UNIVERSITY OF TECHNOLOGY

 

IMAGE: THE TOMATO PICKER ROBOT DESIGNED WITH CHATGPT BY RESEARCHERS FROM TU DELFT AND EPFL MOVES THROUGH A TESTING ENVIRONMENT view more 

CREDIT: © ADRIEN BUTTIER / EPFL

Poems, essays and even books – is there anything the open AI platform ChatGPT can’t handle? These new AI developments have inspired researchers at TU Delft and the Swiss technical university EPFL to dig a little deeper: For instance, can ChatGPT also design a robot? And is this a good thing for the design process, or are there risks? The researchers published their findings in Nature Machine Intelligence.

What are the greatest future challenges for humanity? This was the first question that Cosimo Della Santina, assistant professor, and PhD student Francesco Stella, both from TU Delft, and Josie Hughes from EPFL, asked ChatGPT. “We wanted ChatGPT to design not just a robot, but one that is actually useful,” says Della Santina. In the end, they chose food supply as their challenge, and as they chatted with ChatGPT, they came up with the idea of creating a tomato-harvesting robot.

Helpful suggestions
The researchers followed all of ChatGPT’s design decisions. The input proved particularly valuable in the conceptual phase, according to Stella. “ChatGPT extends the designer's knowledge to other areas of expertise. For example, the chat robot taught us which crop would be most economically valuable to automate.” But ChatGPT also came up with useful suggestions during the implementation phase: “Make the gripper out of silicone or rubber to avoid crushing tomatoes” and “a Dynamixel motor is the best way to drive the robot”. The result of this partnership between humans and AI is a robotic arm that can harvest tomatoes.

ChatGPT as a researcher
The researchers found the collaborative design process to be positive and enriching. “However, we did find that our role as engineers shifted towards performing more technical tasks,” says Stella. In Nature Machine Intelligence, the researchers explore the varying degrees of cooperation between humans and Large Language Models (LLM), of which ChatGPT is one. In the most extreme scenario, AI provides all the input to the robot design, and the human blindly follows it. In this case, the LLM acts as the researcher and engineer, while the human acts as the manager, in charge of specifying the design objectives.

Risk of misinformation
Such an extreme scenario is not yet possible with today’s LLMs. And the question is whether it is desirable. “In fact, LLM output can be misleading if it is not verified or validated. AI bots are designed to generate the ‘most probable’ answer to a question, so there is a risk of misinformation and bias in the robotic field,” Della Santina says. Working with LLMs also raises other important issues, such as plagiarism, traceability and intellectual property.

Della Santina, Stella and Hughes will continue to use the tomato-harvesting robot in their research on robotics. They are also continuing their study of LLMs to design new robots. Specifically, they are looking at the autonomy of AIs in designing their own bodies. “Ultimately an open question for the future of our field is how LLMs can be used to assist robot developers without limiting the creativity and innovation needed for robotics to rise to the challenges of the 21st century,” Stella concludes.

A tomato picker robot designed by ChatGPT and researchers from TU Delft and EPFL in a field test together with a researcher

A robot tomato picker arm designed by ChatGPT and researchers from TU Delft and EPFL "looks' at the camera

CREDIT

© Adrien Buttier / EPFL

JOURNAL

Nature Machine Intelligence

DOI

10.1038/s42256-023-00669-7 

METHOD OF RESEARCH

Experimental study

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

How can LLMs transform the robotic design process?

ARTICLE PUBLICATION DATE

7-Jun-2023

Computational model mimics humans’ ability to predict emotions 

Using insights into how people intuit others’ emotions, researchers have designed a model that approximates this aspect of human social intelligence.

Peer-Reviewed Publication

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

CAMBRIDGE, MA -- When interacting with another person, you likely spend part of your time trying to anticipate how they will feel about what you’re saying or doing. This task requires a cognitive skill called theory of mind, which helps us to infer other people’s beliefs, desires, intentions, and emotions.

MIT neuroscientists have now designed a computational model that can predict other people’s emotions — including joy, gratitude, confusion, regret, and embarrassment — approximating human observers’ social intelligence. The model was designed to predict the emotions of people involved in a situation based on the prisoner’s dilemma, a classic game theory scenario in which two people must decide whether to cooperate with their partner or betray them. 

To build the model, the researchers incorporated several factors that have been hypothesized to influence people’s emotional reactions, including that person’s desires, their expectations in a particular situation, and whether anyone was watching their actions.

“These are very common, basic intuitions, and what we said is, we can take that very basic grammar and make a model that will learn to predict emotions from those features,” says Rebecca Saxe, the John W. Jarve Professor of Brain and Cognitive Sciences, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the study.

Sean Dae Houlihan PhD ’22, a postdoc at the Neukom Institute for Computational Science at Dartmouth College, is the lead author of the paper, which appears in Philosophical Transactions A. Other authors include Max Kleiman-Weiner PhD ’18, a postdoc at MIT and Harvard University; Luke Hewitt PhD ’22, a visiting scholar at Stanford University; and Joshua Tenenbaum, a professor of computational cognitive science at MIT and a member of the Center for Brains, Minds, and Machines and MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).

Predicting emotions

While a great deal of research has gone into training computer models to infer someone’s emotional state based on their facial expression, that is not the most important aspect of human emotional intelligence, Saxe says. Much more important is the ability to predict someone’s emotional response to events before they occur.

“The most important thing about what it is to understand other people's emotions is to anticipate what other people will feel before the thing has happened,” she says. “If all of our emotional intelligence was reactive, that would be a catastrophe.”

To try to model how human observers make these predictions, the researchers used scenarios taken from a British game show called “Golden Balls.” On the show, contestants are paired up with a pot of $100,000 at stake. After negotiating with their partner, each contestant decides, secretly, whether to split the pool or try to steal it. If both decide to split, they each receive $50,000. If one splits and one steals, the stealer gets the entire pot. If both try to steal, no one gets anything.

Depending on the outcome, contestants may experience a range of emotions — joy and relief if both contestants split, surprise and fury if one’s opponent steals the pot, and perhaps guilt mingled with excitement if one successfully steals.

To create a computational model that can predict these emotions, the researchers designed three separate modules. The first module is trained to infer a person’s preferences and beliefs based on their action, through a process called inverse planning.

“This is an idea that says if you see just a little bit of somebody's behavior, you can probabilistically infer things about what they wanted and expected in that situation,” Saxe says.

Using this approach, the first module can predict contestants’ motivations based on their actions in the game. For example, if someone decides to split in an attempt to share the pot, it can be inferred that they also expected the other person to split. If someone decides to steal, they may have expected the other person to steal, and didn’t want to be cheated. Or, they may have expected the other person to split and decided to try to take advantage of them.

The model can also integrate knowledge about specific players, such as the contestant’s occupation, to help it infer the players’ most likely motivation.

The second module compares the outcome of the game with what each player wanted and expected to happen. Then, a third module predicts what emotions the contestants may be feeling, based on the outcome and what was known about their expectations. This third module was trained to predict emotions based on predictions from human observers about how contestants would feel after a particular outcome. The authors emphasize that this is a model of human social intelligence, designed to mimic how observers causally reason about each other’s emotions, not a model of how people actually feel.

“From the data, the model learns that what it means, for example, to feel a lot of joy in this situation, is to get what you wanted, to do it by being fair, and to do it without taking advantage,” Saxe says.

Core intuitions

Once the three modules were up and running, the researchers used them on a new dataset from the game show to determine how the models’ emotion predictions compared with the predictions made by human observers. This model performed much better at that task than any previous model of emotion prediction.

The model’s success stems from its incorporation of key factors that the human brain also uses when predicting how someone else will react to a given situation, Saxe says. Those include computations of how a person will evaluate and emotionally react to a situation, based on their desires and expectations, which relate to not only material gain but also how they are viewed by others.

“Our model has those core intuitions, that the mental states underlying emotion are about what you wanted, what you expected, what happened, and who saw. And what people want is not just stuff. They don’t just want money; they want to be fair, but also not to be the sucker, not to be cheated,” she says.

In future work, the researchers hope to adapt the model so that it can perform more general predictions based on situations other than the game-show scenario used in this study. They are also working on creating models that can predict what happened in the game based solely on the expression on the faces of the contestants after the results were announced.

###

The research was funded by the McGovern Institute; the Paul E. and Lilah Newton Brain Science Award; the Center for Brains, Minds, and Machines; the MIT-IBM Watson AI Lab; and the Multidisciplinary University Research Initiative.

---

JOURNAL

Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences

DOI

10.1098/rsta.2022.0047 

METHOD OF RESEARCH

Computational simulation/modeling

ARTICLE TITLE

Emotion prediction as computation over a generative theory of mind

ARTICLE PUBLICATION DATE

5-Jun-2023

New AI boosts teamwork training

Peer-Reviewed Publication

NORTH CAROLINA STATE UNIVERSITY

Researchers have developed a new artificial intelligence (AI) framework that is better than previous technologies at analyzing and categorizing dialogue between individuals, with the goal of improving team training technologies. The framework will enable training technologies to better understand how well individuals are coordinating with one another and working as part of a team.

“There is a great deal of interest in developing AI-powered training technologies that can understand teamwork dynamics and modify their training to foster improved collaboration among team members,” says Wookhee Min, co-author of a paper on the work and a research scientist at North Carolina State University. “However, previous AI architectures have struggled to accurately assess the content of what team members are sharing with each other when they communicate.”

“We’ve developed a new framework that significantly improves the ability of AI to analyze communication between team members,” says Jay Pande, first author of the paper and a Ph.D. student at NC State. “This is a significant step forward for the development of adaptive training technologies that aim to facilitate effective team communication and collaboration.”

The new AI framework builds on a powerful deep learning model that was trained on a large, text-based language dataset. This model, called the Text-to-Text Transfer Transformer (T5), was then customized using data collected during squad-level training exercises conducted by the U.S. Army.

“We modified the T5 model to use contextual features of the team – such as the speaker’s role – to more accurately analyze team communication,” Min says. “That context can be important. For example, something a team leader says may need to be viewed differently than something another team member says.”

To test the performance of the new framework, the researchers compared it to two previous AI technologies. Specifically, the researchers tested the ability of all three AI technologies to understand the dialogue within a squad of six soldiers during a training exercise.

The AI framework was tasked with two things: classify what sort of dialogue was taking place, and follow the flow of information within the squad. Classifying the dialogue refers to determining the purpose of what was being said. For example, was someone requesting information, providing information, or issuing a command? Following the flow of information refers to how information was being shared within the team. For example, was information being passed up or down the chain of command?

“We found that the new framework performed substantially better than the previous AI technologies,” Pande says.

“One of the things that was particularly promising was that we trained our framework using data from one training mission, but tested the model’s performance using data from a different training mission,” Min says. “And the boost in performance over the previous AI models was notable – even though we were testing the model in a new set of circumstances.”

The researchers also note that they were able to achieve these results using a relatively small version of the T5 model. That’s important, because it means that they can get analysis in fractions of a second without a supercomputer.

“One next step for this work includes exploring the extent to which the new framework can be applied to a variety of other training scenarios,” Pande says.

“We tested the new framework with training data that was transcribed from audio files into text by humans,” Min says. “Another next step will involve integrating the framework with an AI model that transcribes audio data into text, so that we can assess the ability of this technology to analyze team communication data in real time. This will likely involve improving the framework’s ability to deal with noises and errors as the AI transcribes audio data.”

The paper, “Robust Team Communication Analytics with Transformer-Based Dialogue Modeling,” will be presented at the 24th International Conference on Artificial Intelligence in Education (AIED 2023), which will be held July 3-7 in Tokyo, Japan. The paper was co-authored by Jason Saville, a former graduate student at NC State; James Lester, the Goodnight Distinguished University Professor in Artificial Intelligence and Machine Learning at NC State; and Randall Spain of the U.S. Army Combat Capabilities Development Command (DEVCOM). Soldier Center.

This research was sponsored by the U.S. Army DEVCOM, Soldier Center under cooperative agreement W912CG-19-2-0001.

---

METHOD OF RESEARCH

Computational simulation/modeling

SUBJECT OF RESEARCH

People

Digital Science boosts pharma industry support following OntoChem acquisition

In a sea of information, AI-based solutions will help to focus industry’s R&D

Business Announcement

DIGITAL SCIENCE

Digital Science is positioning itself to play an even greater role in the pharmaceutical industry’s all-important drug discovery, by helping industry sift through a sea of information and focus on the research that matters.

To achieve this, Digital Science – a technology company serving stakeholders right across the research ecosystem – has fully acquired OntoChem GmbH, a company highly specialized in AI-based solutions for finding and extracting key information from internal and external data and text, especially published research.

OntoChem has become the newest member of the Digital Science family following an almost two-year partnership between the two companies. OntoChem will continue to work as part of Digital Science’s portfolio product Dimensions – the world’s largest linked research database and data infrastructure provider – which will bring strategic value to the pharmaceutical industry as well as to the wider health and medical sector.

Based in Halle, Germany, OntoChem has 18 years’ experience creating innovative technologies for industry. OntoChem’s text mining, natural language processing and semantic data extraction technologies enable companies to obtain information from both unstructured and structured data, turning it into new knowledge for research and discovery, as well as for strategic decision-making. These tools are utilized by pharmaceutical, material science and technology-driven businesses.

OntoChem currently indexes more than 600 million public documents with 30 million semantic ontology concepts and 150 million synonyms, in fields such as compounds, proteins, diseases, drugs, materials, methods, devices, species, and many more. In compounds alone, OntoChem accesses data from research into 2 billion compounds, which are critical to drug discovery.

Lutz Weber, CEO OntoChem, said: “More and more, pharmaceutical companies are rapidly advancing their research with the use of AI, machine learning, and other technologies, to accelerate their discoveries, and to translate those discoveries into real outcomes.

“One of the biggest issues for pharmaceutical companies is the ‘data dilemma’ – there is so much information to sift through that it can be hard to know where to look or how to focus. Even in one field, such as cancer or diabetes, there is a sea of new knowledge being generated each day in very specific areas of research. This is where our work can help to provide that focus, assisting companies with their discovery and decision-making.

“At OntoChem, we are delighted to become part of Digital Science – this will greatly support our long-term vision to extract the world-wide and comprehensive semantic knowledge from any scientific and related documents or databases, both from enterprise as well as public sources.”

Christian Herzog, Chief Product Officer for Digital Science and co-Founder of Dimensions, said: “OntoChem is a strong strategic fit for Digital Science in combination with our flagship product, Dimensions. The pharmaceutical industry will directly benefit from this acquisition, and through their discoveries our work will benefit the health of individuals and communities right around the world.

“OntoChem’s highly specialized search and extraction tools combined with Dimensions’ access to data from hundreds of millions of research papers, linked grants, patents and clinical trials, and access to data via Google BigQuery, creates a powerful analytical environment. We are already working with life sciences companies, including those in drug research and development, and we expect this to continue and grow in the future.

“In addition to industry, research institutions, governments and funding bodies will greatly benefit from these unique insights suited to discovery and decision-making.”

 

About Digital Science

Digital Science is a technology company working to make research more efficient. We invest in, nurture and support innovative businesses and technologies that make all parts of the research process more open and effective. Our portfolio includes admired brands including Altmetric, Dimensions, Figshare, ReadCube, Symplectic, IFI CLAIMS Patent Services, Overleaf, Ripeta, Writefull, and metaphacts. We believe that together, we can help researchers make a difference. Visit www.digital-science.com and follow @digitalsci on Twitter or on LinkedIn.

About Dimensions

Part of Digital Science, Dimensions is the largest linked research database and data infrastructure provider, re-imagining research discovery with access to grants, publications, clinical trials, patents and policy documents all in one place. www.dimensions.ai

About OntoChem GmbH

OntoChem offers high-performance and highly customizable text analysis and data mining products that can be tailored to meet the specific needs of every client. OntoChem’s main products are: SciWalker – a semantic search and analysis solution to support life and material science communities; OntoChem Processor – a semantic processor tool where clients can enter text from any source; OntoChem Ontologies – a range of ontologies for specific medical and biochemical applications.

OntoChem's information discovery tools are used by small and large life and material science companies to find information by automatically indexing and analyzing internal as well as external data collections. This data provides the raw material for AI and machine learning methods for predictive drug discovery or materials design. OntoChem is proud to work with some of the most influential companies in the fields of pharma, chemistry, specialty chemistry, material science, publishing and IT. Visit www.ontochem.com.

 

Media contacts

Simon Linacre, Head of Content, Brand & Press, Digital Science: Mobile: +44 7484 381477, s.linacre@digital-science.com

David Ellis, Press, PR & Social Manager, Digital Science: Mobile +61 447 783 023, d.ellis@digital-science.com

AI-generated academic science writing can be identified with over 99% accuracy

Peer-Reviewed Publication

CELL PRESS

 

IMAGE: CHATGPT VS HUMAN view more 

CREDIT: HEATHER DESAIRE AND ROMANA JAROSOVA, UNIVERSITY OF KANSAS

The debut of artificial intelligence chatbot ChatGPT has set the world abuzz with its ability to churn out human-like text and conversations. Still, many telltale signs can help us distinguish AI chatbots from humans, according to a study published on June 7 in the journal Cell Reports Physical Science. Based on the signs, the researchers developed a tool to identify AI-generated academic science writing with over 99% accuracy.

“We tried hard to create an accessible method so that with little guidance, even high school students could build an AI detector for different types of writing,” says first author Heather Desaire, a professor at the University of Kansas. “There is a need to address AI writing, and people don’t need a computer science degree to contribute to this field.”

“Right now, there are some pretty glaring problems with AI writing," says Desaire. "One of the biggest problems is that it assembles text from many sources and there isn't any kind of accuracy check — it's kind of like the game Two Truths and a Lie."

Although many AI text detectors are available online and perform fairly well, they weren’t built specifically for academic writing. To fill the gap, the team aimed to build a tool with better performance precisely for this purpose. They focused on a type of article called perspectives, which provide an overview of specific research topics written by scientists. The team selected 64 perspectives and created 128 ChatGPT-generated articles on the same research topics to train the model. When they compared the articles, they found an indicator of AI writing — predictability.

Contrary to AI, humans have more complex paragraph structures, varying in the number of sentences and total words per paragraph, as well as fluctuating sentence length. Preferences in punctuation marks and vocabulary are also a giveaway. For example, scientists gravitate towards words like "however," "but" and "although," while ChatGPT often uses "others" and "researchers" in writing. The team tallied 20 characteristics for the model to look out for.

When tested, the model aced a 100% accuracy rate at weeding out AI-generated full perspective articles from those written by humans. For identifying individual paragraphs within the article, the model had an accuracy rate of 92%. The research team's model also outperformed an available AI text detector on the market by a wide margin on similar tests.

Next, the team plans to determine the scope of the model's applicability. They want to test it on more extensive datasets and across different types of academic science writing. As AI chatbots advance and become more sophisticated, the researchers also want to know if their model will stand.

"The first thing people want to know when they hear about the research is 'Can I use this to tell if my students actually wrote their paper?'" said Desaire. While the model is highly skilled at distinguishing between AI and scientists, Desaire says it was not designed to catch AI-generated student essays for educators. However, she notes that people can easily replicate their methods to build models for their own purposes.

###

Cell Reports Physical Science, Desaire et al. “Distinguishing academic science writing from humans or ChatGPT with over 99% accuracy using off-the-shelf machine learning tools” https://www.cell.com/cell-reports-physical-science/fulltext/S2666-3864(23)00200-X

Cell Reports Physical Science (@CellRepPhysSci), published by Cell Press, is a new broad-scope, open access journal that publishes cutting-edge research across the spectrum of the physical sciences, including chemistry, physics, materials science, energy science, engineering, and related interdisciplinary work. Visit https://www.cell.com/cell-reports-physical-science/home. To receive Cell Press media alerts, please contact press@cell.com.

---

JOURNAL

Cell Reports Physical Science

DOI

10.1016/j.xcrp.2023.101426 

METHOD OF RESEARCH

Computational simulation/modeling

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

Distinguishing academic science writing from humans or ChatGPT with over 99% accuracy using off-the-shelf machine learning tools.

ARTICLE PUBLICATION DATE

7-Jun-2023

New “AI doctor” predicts hospital readmission and other health outcomes

Tool designed to smooth hospital operations for better patient care

Peer-Reviewed Publication

NYU LANGONE HEALTH / NYU GROSSMAN SCHOOL OF MEDICINE

An artificial intelligence (AI) computer program can read physicians’ notes to accurately estimate patients’ risk of death, length of hospital stay, and other factors important to care. Designed by a team led by researchers at NYU Grossman School of Medicine, the tool is currently in use in its affiliated hospitals to predict the chances that a patient who is discharged will be readmitted within a month.

Experts have long explored computer algorithms meant to improve healthcare, with some having been shown to make valuable clinical predictions. However, few are in use because computers best process information laid out in neat tables, while physicians typically write in creative, individualized language that reflects how humans think.

Cumbersome data reorganization has been an obstacle, researchers say, but a new type of AI, large language models (LLM), can “learn” from text without needing specially formatted data.

In a study publishing online June 7 in the journal Nature, the research team designed an LLM called NYUTron that can be trained using unaltered text from electronic health records to make useful assessments about patient health status. The results revealed that the program could predict 80% of those who were readmitted, a roughly 5% improvement over a standard, non-LLM computer model that required reformatting of medical data.

“Our findings highlight the potential for using large language models to guide physicians about patient care,” said study lead author Lavender Jiang, BSc, a doctoral student at NYU’s Center for Data Science. “Programs like NYUTron can alert healthcare providers in real time about factors that might lead to readmission and other concerns so they can be swiftly addressed or even averted.”

Jiang adds that by automating basic tasks, the technology may speed up workflow and allow physicians to spend more time speaking with their patients.

Large language models use specialized computer algorithms to predict the best word to fill in a sentence based on how likely real people would use a particular term in that context. The more data used to “teach” the computer how to recognize such word patterns, the more accurate its guesses become over time, adds Jiang.

For their study, the researchers trained NYUTron using millions of clinical notes collected from the electronic health records of 336,000 men and women who had received care within the NYU Langone hospital system between January 2011 and May 2020. The resulting 4.1-billion-word language “cloud” included any record written by a doctor, such as radiology reports, patient progress notes, and discharge instructions. Notably, language was not standardized among physicians, and the program could even interpret abbreviations unique to a particular writer.

According to the findings, NYUTron identified 85% of those who died in the hospital (a 7% improvement over standard methods) and estimated 79% of patients’ actual length of stay (a 12% improvement over the standard model). The tool also successfully assessed the likelihood of additional conditions accompanying a primary disease (comorbidity index) as well as the chances of an insurance denial.

“These results demonstrate that large language models make the development of ‘smart hospitals’ not only a possibility, but a reality,” said study senior author and neurosurgeon Eric Oermann, MD. “Since NYUTron reads information taken directly from the electronic health record, its predictive models can be easily built and quickly implemented through the healthcare system.”

Oermann, an assistant professor in the Departments of Neurosurgery and Radiology at NYU Langone Health, adds that future studies may explore the model’s ability to extract billing codes, predict risk of infection, and identify the right medication to order, among other potential applications.

He cautions that NYUTron is a support tool for healthcare providers and is not intended as a replacement for provider judgment tailored to individual patients.

Funding for the study was provided by National Institutes of Health grants P30CA016087 and R01CA226527. Further funding was provided by the W.M. Keck Foundation Medical Research Grant.

Oermann is a consultant for Sofinnova Partners, a venture-capital firm that specializes in the development of biotechnologies, sustainability, and pharmaceuticals. He also has equity in Artisight Inc., a company that produces a software for healthcare organization operations, and his spouse is employed by Mirati Therapeutics, which develops cancer therapies. The terms and conditions of these arrangements are being managed in accordance with the policies and practices of NYU Langone Health.

In addition to Jiang and Oermann, other NYU investigators involved in the study were Xujin Chris Liu, BS;  Mustafa Nasir-Moin, BA; Duo Wang, PhD;  Anas Abidin, PhD; Kevin Eaton, MD; Howard Riina, MD;  Ilya Laufer, MD; Paawan Punjabi, MD; Madeline Miceli, MD; Nora Kim, MD; Cordelia Orillac, MD; Zane Schnurman, MD; Christopher Livia, MD, PhD; Hannah Weiss, MD; David Kurland, MD, PhD; Sean Neifert, MD; Yosef Dastagirzada, MD; Douglas Kondziolka, MD; Alexander Cheung, MA; Grace Yang; Ming Cao; Monda Flores; Anthony Costa, PhD; Yindalon Aphinyanaphongs, MD, PhD; and Kygunghyun Cho, PhD. Other study authors included Nima Pour Nejatian, PhD, MBA, at nVidia in Santa Clara, Calif., whose computer hardware was used to build NYUTron.

---

JOURNAL

Nature

DOI

10.1038/s41586-023-06160-y 

SUBJECT OF RESEARCH

People

ARTICLE TITLE

Health system-scale language models are all-purpose prediction engines

ARTICLE PUBLICATION DATE

7-Jun-2023

COI STATEMENT

Oermann is a consultant for Sofinnova Partners, a venture-capital firm that specializes in the development of biotechnologies, sustainability, and pharmaceuticals. He also has equity in Artisight Inc., a company that produces a software for healthcare organization operations, and his spouse is employed by Mirati Therapeutics, which develops cancer therapies. The terms and conditions of these arrangements are being managed in accordance with the policies and practices of NYU Langone Health.