It’s possible that I shall make an ass of myself. But in that case one can always get out of it with a little dialectic. I have, of course, so worded my proposition as to be right either way (K.Marx, Letter to F.Engels on the Indian Mutiny)
Wednesday, November 18, 2026
Substance use, mental health, and respiratory infections shifting burden of disease in Canada
Analysis of Global Burden of Disease Study of 204 countries looks at Canadian trends
The conditions driving the burden of disease in Canada are shifting, according to an analysis of the Global Burden of Disease (GBD) Study published in CMAJ (Canadian Medical Association Journal) https://www.cmaj.ca/lookup/doi/10.1503/cmaj.260996. Cancers, cardiovascular diseases, musculoskeletal disorders, and unintentional injuries were the four leading causes of health loss in Canada in 2023, accounting for half of all health loss. But the fastest growing contributors between 2010 and 2023 were substance use disorders, mental health conditions, and respiratory infections, including COVID-19, after accounting for age.
The GBD Study, led by the Institute for Health Metrics and Evaluation at the University of Washington, estimates health loss from diseases, injuries, and risk factors in 204 countries and territories. It covers 292 causes of death, 375 diseases and injuries, and 88 modifiable risk factors. A team led by Canadian researchers analyzed these estimates to describe health loss in Canada in 2023, trends from 2010 to 2023, and how Canada compares with other G7 countries.
To measure health loss, the study uses disability-adjusted life years (DALYs). One DALY represents one year of healthy life lost, whether due to premature death or to disability from illness or injury. The higher the number, the greater the health loss.
“The point of measuring burden this way is to see where health loss is growing, not just where it is largest. Cancer and cardiovascular disease remain the leading causes of health loss, but they are declining once you account for an aging population. Mental health and substance use disorders are moving the other way,” says Dr. Heather Orpana, senior research scientist at the Public Health Agency of Canada and adjunct professor at the University of Ottawa, Ottawa, Ontario. “Unlike many of the leading causes of health loss, mental health and substance use disorders tend to affect people earlier in life, during the years when they are finishing school and starting work. That means the impact goes well beyond what we measure in a health study.”
Highlights:
Noncommunicable diseases were responsible for the vast majority of all DALYs, making up 82% of the disease burden in Canada in 2023.
Cancer (neoplasms) was the leading cause of all-age DALY rates in both males and females.
Mental disorders were the fifth leading cause of health loss for both sexes combined.
If mental disorders, substance use disorders, and self-harm were combined, this group would be the second leading cause of health loss, behind cancer.
Health loss due to substance use disorders had the largest percentage increase (90%) between 2010 and 2023.
There were also large increases (67%) in health loss from respiratory infections, including COVID-19 and tuberculosis, from 2010 to 2023. COVID-19, absent before 2020, was responsible for most of the increase in the all-age DALY rate from 2010 to 2023.
For females, 57% of health loss came from years lived with disability; for males, 55% came from premature death. About 40% of total health loss as measured by DALYs could be attributable to risk factors that were measured in the GBD study.
Behavioural risk factors such as tobacco use or low physical activity accounted for about a quarter of all-age DALYs, and metabolic risk factors such as high cholesterol or high blood sugar accounted for almost 20%.
Environmental and occupational risks, like lead exposure or air pollution accounted for 6%.
Totals don’t add up due to overlapping risks and mediating pathways.
Tobacco use continued to be the leading risk factor contributing to all-age DALY rates in 2023. Although there was no significant change in all-age attributable DALYs due to tobacco use between 2010 and 2023, after accounting for the aging population, the attributable DALY rate due to tobacco use decreased by almost 20%.
All-age DALYs attributable to drug use as a risk factor increased by 118% between 2010 and 2023.
“Almost 40% of the health loss in Canada can be traced to risk factors we already work hard to address: things like tobacco, diet, blood pressure, body weight, and physical inactivity. That portion of the burden is, in principle, preventable,” says coauthor Dr. Justin Lang, research scientist at the Public Health Agency of Canada. “Many of the risk factors we measure, like diet or physical inactivity, are shaped by conditions people don’t fully control. Focusing only on individual behaviour can draw attention away from those upstream causes.”
Canada’s overall burden of disease was comparable to that of other G7 countries, while Canada’s life expectancy fell in the middle. The United States had the highest age-standardized DALY rate in G7 countries, and Japan and Italy had the lowest. Health loss due to substance use disorders was ranked substantially higher in Canada and the US, and ranks for cardiovascular disease and cancer (neoplasms) were similar among G7 countries.
Oakland, CA—Among people in inpatient alcohol treatment, the share attending Alcoholics Anonymous or similar groups at least weekly fell from more than half to fewer than four in 10 in just seven years, according to a study published in the Journal of Studies on Alcohol and Drugs. Latinx/Hispanic, Black and American Indian/Alaska Native patients were less likely than White patients to attend weekly, even after accounting for factors such as age, employment and alcohol use.
Researchers analyzed federal data on nearly 2.8 million discharges from publicly funded substance use treatment facilities in every U.S. state except Oregon between 2015 and 2022. The study includes adults treated for alcohol use, either alone or alongside other drug use. Analyses examined inpatients and outpatients separately.
“Mutual-help groups are widely available, low-cost and one of the most effective resources we have for supporting long-term recovery from alcohol use disorder,” said Sarah E. Zemore, Ph.D., lead author and senior scientist at the Alcohol Research Group, a program of the Public Health Institute. “Attendance has dropped nationally, and it is considerably lower for several racial and ethnic groups. Treatment programs are well placed to change that by actively connecting patients with meetings where they feel they belong.”
A steep national decline
Weekly attendance, defined as attending four or more meetings in the 30 days before discharge, dropped substantially over the study period:
Inpatient treatment: from 53.9% in 2015 to 39.2% in 2022
Outpatient treatment: from 19.7% in 2015 to 16.0% in 2021 and 2022
After accounting for other factors, the odds of weekly attendance in 2022, relative to 2015, were about one-third lower for outpatients and about one-half lower for inpatients. The sharpest drops came in 2020, when the COVID-19 pandemic limited in-person meetings. For example, inpatient attendance fell from 49.8% in 2019 to 41.7% in 2020 and remained low through 2022.
The researchers suggest that a shift in the treatment field away from abstinence-focused goals may also have played a role, along with wider use of medications for substance use, which can be stigmatized in some 12-step programs.
“Medications and mutual-help groups aren’t competing options. Many people do best with both, so programs should encourage support group participation whether or not a person is taking medication,” said Katherine Karriker-Jaffe, Ph.D., co-author and senior director at RTI International, an independent scientific research institute.
Gaps across racial and ethnic groups
Compared with White patients, odds of weekly attendance across settings were:
36% to 39% lower for Latinx/Hispanic patients
28% lower for Black patients
22% to 28% lower for American Indian/Alaska Native patients
10% to 11% lower for patients of other races and ethnicities
In contrast, Asian/Pacific Islander patients had 60% to 66% higher odds of weekly attendance than White patients. Women had slightly higher odds of weekly attendance than men. Compared with adults ages 45-54, young adults (18–24) and adults 65 and older had lower odds of weekly attendance, as did those who were married (vs. unmarried), employed (vs. unemployed) and those without a college degree.
Why it matters
Rigorous research shows that participation in AA and similar groups improves alcohol outcomes. The decline in mutual-help group participation therefore likely means poorer treatment outcomes for many patients.
The study can’t show why attendance differs between racial and ethnic groups. Earlier research suggests some people avoid groups they see as designed for White men, or where they have faced mistrust or discrimination. But studies finding higher participation among Black clients in some urban areas suggest the key barrier may be less the groups themselves than being a demographic minority at the meetings available to them.
The researchers caution that simply expanding standard programs that encourage 12-step participation may not close these gaps, because such programs have not proven especially effective for racial and ethnic minority participants. They call for research into the barriers these groups face and for culturally tailored approaches, such as more Spanish-language meetings, referrals to culturally adapted groups and connections to peer support specialists and recovery community centers.
About the data
The data come from the Substance Abuse and Mental Health Services Administration’s Treatment Episode Data Set–Discharges (TEDS-D). They cover publicly funded facilities (about half of all U.S. substance use treatment facilities) and exclude private and most federal programs. The data count treatment episodes, not individuals, but the main findings held when analyses were limited to first-time patients.
Study:
Zemore, SE, Rosen, EM, Mulia, N, Rodriguez Borja, I, Karriker-Jaffe, KJ. Race/ethnicity and Other Sociodemographic Correlates of Weekly Mutual-Help Group Attendance During Alcohol Treatment in 2015–2022 National Data. J Stud Alcohol Drugs. 2026 November. https://doi.org/10.15288/jsad.26-00104.
Funding: This research was supported by the NIH National Institute on Alcohol Abuse and Alcoholism (R01AA029812, R01AA027920).
###
The Alcohol Research Group (ARG), a program of the Public Health Institute, is a nonprofit research group with more than 65 years of experience studying alcohol and other drug use and their consequences for individuals, families and communities. Based in Berkeley, California, ARG's multidisciplinary team of scientists, methodologists and public health researchers works to generate evidence that drives meaningful change in policy, services, and public understanding. For more information, visit www.arg.org.
The Public Health Institute is a national independent nonprofit that advances the health and wellbeing of all people and communities, particularly those facing the greatest barriers. Rooted in 60 years of public health experience, PHI moves ideas into action through practice, policy, research and community engagement, while serving as the operational backbone for more than 40 programs delivering evidence-based solutions to complex health challenges. For more information, visit www.phi.org.
RTI International is an independent scientific research institute dedicated to improving the human condition. Our vision is to address the world's most critical problems with technical and science-based solutions in pursuit of a better future. Clients rely on us to answer questions that demand an objective and multidisciplinary approach—one that integrates expertise across social, statistical, data, and laboratory sciences, engineering, and other technical disciplines to solve the world’s most challenging problems. For more information, visit www.rti.org.
AMHERST, Mass. — Researchers at the University of Massachusetts Amherst have discovered how, during a bank-to-bank transfer, money can be deposited into one bank account without the funds ever leaving the second bank. This new class of cyberthreat is called Credit-Without-Debit (CWD) attacks and the researchers have demonstrated three ways hackers can use it to steal money.
“Our goal is to strengthen financial security,” says Taqi Raza, assistant professor in the Riccio College of Engineering at UMass Amherst. “The way we strengthen it is we find the loopholes and tell the banks, ‘okay, fix this problem.’”
Unlike previous research by this group, which focused on credit card and digital wallet vulnerabilities, this new study examines debit transactions, or ACH transactions. This system permanently moves money between banks—as opposed to creating a debt—and enables any bank-to-bank transaction, including direct deposit of salaries, bill payments (including credit card bills), mortgage services, and withdrawals from digital wallets and various payment apps.
The way an ACH transaction should work is that transfers are processed through multi-stage interbank clearing and settlement. Consider the situation where a person has bank accounts at two different banks and wants to move money from Bank A into Bank B. The person initiates the transfer from Bank B, which pulls the funds from the account at Bank A.
The transaction is authorized when Bank B recognizes that Bank A has a valid routing and account number and then authorizes the transaction. There is a two-day payment processing window for Bank A to transfer the money, so to expedite access to funds, some banks provisionally post funds in their customer’s account (in this example, Bank B posts these funds). Bank A checks that its customer’s account has sufficient funds, that there is valid authorization, and that there are no account holds, and then sends the money to Bank B. The interbank settlement is complete.
However, in a CWD attack, the attacker finds a victim with a frozen account, meaning the account can receive money but not send it. They link the victim’s account to their own and then pull the victim’s funds into their own account. The attacker’s bank posts the provisional funds. However, because the victim’s bank is frozen, the settlement payment is never made. In the two days it takes for this payment to be processed and rejected, the attacker has withdrawn the provisional funds, resulting in a direct financial loss to the bank. The researchers have created an interactive graphic to demonstrate how this attack works.
“The key root cause of all these attacks is that banks have treated distinct security checkpoints as interchangeable,” explains Raza.
Once they discovered the vulnerabilities, the researchers coordinated with affected banks to resolve the issues, which have not been reproducible since. The team presented their paper on the work today at the ACM Conference on Computer and Communications Security (CCS), 2026.
The researchers found two additional instances where this credit-without-debit logic resulted in successful attacks. One attack used microdeposits from third-party payment apps used to verify bank account ownership; the other paid off a credit card bill using a frozen account.
“We focused on what the user can observe and what they can influence from their banking apps,” says Ali Malik, Ph.D. candidate in electrical and computer science at UMass Amherst. “They can initiate payments, they can cancel those at random moments, they can observe the progress of that transaction as it moves from processing, to settle, to cancel, to rejected. It’s simple stuff. We defined our threat model around that.”
Even with these vulnerabilities discovered, Raza recommends that consumers always follow security best practices, such as enabling two-factor authentication, monitoring bank statements for fraudulent activity and turning on text, email, or banking app notifications whenever a debit or credit happens on an account.
Thousands of deceptive advertisements have slipped through Google's advertising safeguards, exposing internet users to fake virus warnings, misleading software promises and potentially dangerous downloads, according to new research from NYU Tandon and Radboud University in the Netherlands.
Using an AI tool they developed called AdLens, the researchers examined nearly 189,000 versions of software-related advertisements from Google's publicly accessible Ads Transparency Center and identified thousands of deceptive ads that collectively generated more than 100 million impressions across European markets.
Some of the Google ads it identified remained active for years, while others continued running even after being reported to the company.
In one case, Google removed an advertisement linked to a known malware operation but left 41 others directing users to the same website. The findings raise questions about how effectively Google detects harmful advertising and whether its enforcement systems address entire campaigns rather than individual ads.
“Google has systems in place to screen advertisements, but we found thousands of deceptive ads that still reached users, sometimes for years,” said Ritik Roongta, NYU Tandon Computer Science and Engineering (CSE) PhD student and lead author of the study. “Even when we reported ads linked to known malware campaigns, removing one didn't necessarily stop dozens of others promoting the same threat. We shouldn't have to rely entirely on advertising platforms to tell us whether their safeguards are working.”
AdLens identified three major categories of deceptive advertising:
Scareware Ads: Ads with fake warnings include those claiming that users' devices were infected, hacked or compromised generated more than 16 million impressions. The researchers identified 238 such ads within the 189k ads examined.
Misleading software promises: The researchers identified 3,346 ads making false or unrealistic claims about what software could do, including recovering photos deleted years earlier. These ads generated nearly 90 million impressions.
Deceptive designs: The researchers identified 258 ads that concealed the advertiser's identity or what was being promoted, sometimes displaying little more than a button labeled "Continue" or "Access Now." One such ad generated 2.1 million impressions.
One advertisement promising to reveal anyone's GPS location using only a phone number generated 14.5 million impressions and remained active for more than two years.
“These ads are only the tip of the iceberg, as the ad transparency platform hosts more than two billion ads of which only a small portion was analysed,” said Roongta. “The high incident of deception within this small sample itself highlights the magnitude of these scams being propagated to the users.”
The researchers also examined apps promoted through deceptive advertisements. Of 147 identified apps, 34 were no longer available in app stores even though their advertisements remained active.
AdLens uses publicly available AI models (small language models or SLMs) to identify suspicious advertising efficiently. It first searches for language resembling known deceptive tactics, then examines the text and images of likely offenders. By concentrating intensive analysis on a smaller group of ads, the system reduces the computing resources needed for large-scale investigations.
The study builds on earlier research by Roongta and Greenstadt, which found that users of Adblock Plus's "Acceptable Ads" feature encountered 13.6% more problematic advertisements than users browsing without ad blockers. Together, the studies reveal shortcomings in systems intended to improve the online advertising experience.
The researchers have made AdLens publicly available so regulators, consumer protection organizations and independent researchers can investigate advertising practices without relying exclusively on platforms' own assessments. Although this study focused on Google, the technology could be adapted to examine other advertising platforms, including Meta and TikTok.
The researchers caution that ad impressions measure exposure, not confirmed financial losses or malware infections. Nevertheless, their findings demonstrate how deceptive advertising can persist despite existing safeguards and underscore the need for stronger enforcement and independent oversight.
In addition to Roongta and Greenstadt, the authors of the AdLens paper are Marwan Adnan Darwish, Masoud Poorghaffar Aghdam, and Gunes Acar, all of Radboud University. The research was supported by the National Science Foundation and the Netherlands Organisation for Scientific Research.
A third of the global AI supply chain is ‘invisible’ to governments, study finds
Researchers identify 137 supply chain elements for global AI production. They compare these to government documents from dozens of countries and find 33% of the elements, mainly from the “physical middle” of AI production (such as wafers, packaging and networking), are largely absent from top-level AI policy worldwide.
“AI sovereignty” is an impossibility for any one nation, including the US and China, and countries such as Japan and the UK have far more influence over the supply of AI than they realise, say experts. A country-by-country report is available alongside the main report.
As much as a third of the AI supply chain receives no attention from national and international policymaking and regulation, according to a new Cambridge report that reveals a vast “zone of invisibility” in the middle of global AI production.
A University of Cambridge team found vulnerabilities across 14 supply-chain layers, with 137 elements produced or convened in over fifty countries. Many critical “chokepoints” sit outside the AI powerhouses of China and the US.
“The US and China are central to the AI story, but this research shows us the extent to which critical technologies and materials across the AI stack have dependencies that extend far beyond their borders,” said Verity Harding, Director of the AI and Geopolitics project. “It shows that no one country can control the AI supply chain alone.”
One Japanese firm has a 95% market share of a material needed for AI chips, the report shows. Even notorious chokepoints depend on other nations: the Dutch firm dominating EUV lithography (ASML) relies on essential parts controlled by German and US companies.
In fact, the study suggests that countries such as Japan, Germany, and some so-called middle powers – from Belgium to South Korea – have far more leverage over global AI production than even their own governments realise.
Researchers describe an “unexpected mismatch” between a nation’s top AI policy documents and its actual positions in the supply chain, calling it a “universal pattern”.
“Most policy discussions reduce artificial intelligence to just a few inputs, such as compute, data and models. This is not an accurate understanding of the complexity of the AI supply chain,” said lead author Dr Siyi Liu.
“Many governments are unaware of the strengths and vulnerabilities they hold at the heart of the global AI supply chain. Understanding this will be the basis of credible bargaining power in AI governance.” she added.
The UK is involved in 41 of the 137 supply-chain elements researchers identify, yet its AI policies name just three, ignoring UK specialisms in chip design and industrial gases. Japan is second only to the US in terms of supply-chain positions, yet its AI policy documents recognise only three.
Cambridge researchers analysed 117 AI-related policy documents from 43 countries and territories, the EU and OECD, and found a fixation on models and datacentres, as well as funding and governance, while the factories, chemicals, chip-design software and packaging technologies vital to the AI supply chain are completely overlooked.
Disruption to any one of these “upstream” components could trigger a domino effect that leads to AI supply shortages, potentially destabilising stock markets, investment and productivity around the world.
Experts from the AI & Geopolitics Project (AIxGEO) at Cambridge’s Bennett School of Public Policy say that total “AI sovereignty”, a nation autonomously operating its own AI supply chain – a goal many nations hint at pursuing – is “impossible”.
The research team created a toolkit for governments to audit their place in the chain. They say it will help policymakers understand where and when to diversify across AI element providers, who they depend on, or when to “exit” entirely from an outside supplier and build national capacity.
For the latest study, the AIxGEO team comprehensively mapped the global AI supply chain to reveal 14 distinct layers, from the raw minerals and the ultra-pure gases needed for semiconductors, to the wafers, chips and servers, and eventually the trained models. Cross-cutting layers include governance standards and shipping routes.
Breaking it down, researchers identified 137 specific “elements”, such as gallium or hyperscale cloud platforms, with a further 351 “sub-elements”: tech specs that determine who can supply it. Nearly 800 firms identified hold supply-chain positions across 56 countries and territories.
However, the team’s analysis of AI policy and regulation documents from dozens of countries as well as intergovernmental organisations shows that just over 33% of the identified supply-chain elements are simply not mentioned.
Most of this “invisibility” is in the “physical middle” of the chain. This includes chip design, packaging and the testing of chips, and some of the hardware and software behind “compute” such as making AI servers and “merchant accelerators” (platforms that allow businesses to adopt AI).
An example of a hidden chokepoint highlighted by the report is the EDA (Electronic Design Automation) software used to design AI chips. It is essential for companies like NVIDIA, AMD and Google.
This software is dominated by three firms – Synopsys, Cadence and Siemens EDA in the US – yet receives almost no policy attention, unlike the chips and LLMs that wouldn’t exist without it.
Another major “blind spot” is advanced packaging: connecting multiple semiconductor chips into high-performance AI processors essential for frontier models. This packaging is overwhelmingly concentrated in Asian countries and territories such as Taiwan, Vietnam and the Philippines.
“Policymakers focus on finished chips rather than the ecosystem behind them. They focus on GPUs but ignore the packaging operations needed to turn those chips into usable AI hardware,” said Liu.
Japan emerges as the world’s most underrated AI power, with the second largest number of supply chain positions, directly ahead of China. The report shows 101 Japanese firms hold positions across 69 elements, from silicon wafers to semiconductor testing. However, Japan’s own policy documents feature only three of these elements.
The researchers point out that a Japanese company, Ajinomoto, controls more than 95% of the global supply of build-up film, an insulating material needed for advanced AI chip manufacturing.
“Japan’s strength comes from equipment, and without it there's no way you can produce the hardware,” said Liu. “The supply chain is sequential, so the trickle down could affect computing capacities of the top frontier models.”
The UK has little AI-related manufacturing, but holds leverage in chip design IP through Cambridge company Arm (even with majority Japanese SoftBank ownership). It is a base for research and expertise, particularly in AI security, and home to valuable health and genomic datasets.
The report highlights absences from reviewed UK policy where it has footholds in the supply chain, such as chip design, and the pure gases needed for semiconductor plants via Linde, one of the world's largest industrial gas companies.
Some of the smaller nations with more clout than they might realise include Belgium, which is strong in recycled semiconductor chemicals and has the major nanoelectronics R&D centre Interuniversity Microelectronics Centre (imec).
Ireland has no production supply-chain position, yet it can affect AI capacity through energy and planning as it provides a European base for big US tech e.g. Google and AMD, making it a “host and infrastructure state”. While Qatar produces much of the world’s helium, a vital supply-chain raw material, and is also a huge potential AI investor.
Earlier this year, Nvidia’s former-CEO Jensen Huang described the AI supply chain as a “five-layer cake” (energy, chips, infrastructure, models, and applications).
Verity Harding added: “The AI supply chain is far more complex, and power over it far more diffused, than some simplified narratives might have us believe. Many governments hold more cards than they realise.”
Notes:
The 14 supply chain layers identified by the report: minerals that must be scooped and refined; ultra-pure gases and chemicals required by a fabrication plant or facility; wafers; machines that print the circuits; chips; memory and packaging around chips; servers; data centres; power and cooling that run data centres; cloud platforms; data; models.
About the AI & Geopolitics project (AIxGEO)
AIxGEO, led by Verity Harding, exists to interrogate the framing of AI development as a zero-sum arms race between two superpowers, and to shift the policy debate towards a more cooperative future so the technology can deliver the broad benefits it promises.
About the Bennett School of Public Policy
The Bennett School is a leading school of public policy in the UK and wider world, for public policy thinking, research, teaching and engagement with practitioners from policy, business and civic organisations. Drawing on Cambridge’s broad talent base, it supports effective decision-making through interdisciplinary research, training, and collaboration.
Pulling on an AI model reveals how it represents physical relationships, and how we can train better AI models that understand physics from first principles
Dongguan Institute of Materials Science and Technology, CAS
A spring illustrates the mechanics-inspired idea: a system’s behavior is revealed by observing how it responds to a controlled change. In the language model, paired scientific questions produce changes in internal representations, depicted as interconnected networks. Targeted interventions then test how selected representations influence the model’s answer preferences. The figure is a conceptual illustration, not a visualization of measured data. Image credit: Markus J. Buehler/MIT.
Large language models are increasingly used to explore scientific knowledge, analyze materials and generate ideas. Yet a correct answer alone does not reveal how a model arrived at it. The answer might reflect a physical relationship, a familiar phrase or a statistical shortcut.
Physical laws describe relationships between quantities. An increase in one quantity can increase a material's response under one law and decrease it under another. Testing whether an AI model represents that distinction requires more than examining its response to a single question.
The Solution
Markus J. Buehler, the Jerry McAfee Professor of Engineering at MIT, investigated this problem using three open-weight models from Google's Gemma family. Inspired by mechanics, he compared matched questions in which a physical condition changed while the surrounding context was controlled. He then measured how the model's hidden states, the internal numerical representations it creates while processing a question, moved in response.
Across 60 materials-science laws, the resulting changes tracked the ordering of direct, neutral and inverse relationships with a rank correlation of 0.910. They correctly oriented 39 of 40 directional laws. When equations and prescribed reasoning instructions were removed, the models still oriented 38 of 40 laws correctly at the smallest tested scale and all 40 at the two larger scales.
The study also tested whether scientific concepts could be read from intermediate hidden states when their names were omitted from the prompt. Some concepts remained recoverable, while others did not. This established a distinction between information that can be read from a model and information that is used in a decision.
To test causal use, the author modified selected internal representations without retraining the model or adding verbal hints. In one controlled test, a fixed direction associated with grain size shifted answer preferences appropriately across 12 grain-refinement and grain-coarsening cases. Further interventions across 24 relationships showed selective causal contributions across all three model sizes.
The effects were context-dependent: the same grain-size intervention did not transfer successfully to every new wording or physical regime tested. The experiments therefore distinguish three levels of evidence: readability, representation and causal use.
The Future
The next step is representation-aware scientific AI: training and evaluating models not only on their answers, but also on whether their internal representations respond consistently to changes in physical conditions. This opens a path to testing models against scientific relationships and identifying where they rely on shortcuts.
The present work establishes the measurements and interventions for selected, known relationships in controlled tasks. Extending them to new scientific problems and model architectures is the next challenge, including applications in developing enhanced training rewards for AI models that foster broader and deeper generalizations of physics principles. This can lead to new first-principles based AI models that think and reason from foundational concepts and that may be far more efficient and powerful than current models.
The Impact
By treating an AI model's internal representations as objects that can be measured and experimentally altered, this research brings an approach inspired by the physical sciences to the study of AI itself. It gives researchers a more precise way to ask what a model represents, what it uses and where its scientific reasoning breaks down.
Publication: M.J. Buehler, "Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model," AI for Science2, 045001 (2026). https://doi.org/10.1088/3050-287X/aeb27b
Why do some artworks stick in our minds, while others fade away soon after we view them? A new study published in Psychological Science suggests that it’s partly because the artists themselves can imbue memorability into their works.
Researchers at the University of Chicago tested this ability by conducting a nationwide art contest, asking artists across the country to create works that were intended to be either memorable or forgettable. They found that not only were artists successful in completing this task, but that a deep learning model was most adept at predicting memorability. These findings provide insights into how artists may intuitively understand memory and how technology can predict it.
“Our study suggests that artists can offer valuable insights into memory,” said Yifei Chen, a coauthor of the new study. “They seem to have an intuitive sense of what makes artworks actually stick in people’s minds.”
Chen and her colleagues collected art contest submissions over the course of three months for the study. They received a total of 86 submissions that adhered to the contest’s rules and first tested the memorability of the artworks with a group of online participants. Overall, the researchers found that the works submitted to the “memorable” category were more likely to be remembered, and that “forgettable” works tended to slip the mind. Analyses also showed that simpler artworks, as well as ones with fewer repetitive elements and less symmetry, had more memorability.
The team then chose 20 of both the most memorable and most forgettable artworks from these subsets. They asked the artists to ship their art to Chicago for a month-long exhibit in an in-person gallery. When the exhibition opened, visitors could freely move through the gallery, after which their memory of the artworks was tested. “The same pattern held in the gallery: Artworks that were more memorable online were also better remembered in person,” Chen said.
“It was a very risky experiment for our lab,” noted Wilma Bainbridge, senior author of the paper and Chen’s advisor, as the gallery only had 40 pieces and had an overarching theme of memory. The process of the experiment was also captured in a short documentary.
“But even still, they forgot the forgettable pieces and remembered the memorable ones, which shows how strong this effect really is,” she said. “It works on the order of just a few minutes, even in highly motivated people in this real-world setting.”
Finally, the researchers tested if memory of artworks could be predicted by artists, online laypeople without any art expertise, or a neural network model called ResMem that was developed by researchers to predict image memorability. In the new study, ResMem performed best at distinguishing the most and least memorable artworks, followed by artists’ submission categories, then ratings by laypeople.
“ResMem had only been trained on photographs,” Chen explained, making these new, original works a test of its ability. And yet, it did “much better than any human can do,” she said, highlighting that “what people think is memorable is different from what actually is.”
Although artists weren’t as proficient as AI in predicting memorability, their success in designing for memory suggests they could be valuable scientific collaborators. Studying their creative choices could help researchers understand which visual features support memory and how those features might be deliberately shaped. “This gives us so much potential for tapping into what drives memorability,” Chen said.
Refining models like ResMem could also help artists find tangible ways to make their works stand out or aid exhibit designers with how to arrange galleries. Bainbridge adds that ResMem could also be applied in a broader educational context, such as finding out what kinds of visuals may help children, patients, or other groups learn information that is difficult to remember. “Our preliminary findings are that if you have a memorable picture with all this information, they learn that additional information better, too,” Bainbridge said.
These findings contribute to a growing body of work examining how the properties of images and the real-world contexts they are encountered in can shape what we remember, Chen said. Because memory research has often focused on processes within the viewer, studying the relationships of visuals and memory in everyday settings is, as Chen says, “a direction that the whole field could explore further in the future.”
Researchers put AI's understanding of people to the test
The framework helps scientists determine whether AI can identify meaningful patterns in human experiences and shows where human insight makes a difference
University of Missouri researchers have developed a way to determine whether artificial intelligence is uncovering meaningful insights from people’s words or simply sorting them into categories.
From interviews and focus groups to open-ended survey responses, researchers look for patterns in what people say to better understand their experiences and perspectives. But identifying meaningful themes takes time, and interpreting people’s words requires more than simply grouping similar responses together.
To determine whether AI can identify meaningful patterns without losing important nuances, a Mizzou team spanning journalism, data science and nursing developed a method to evaluate computer-generated topics. The approach examines how closely those topics match themes identified by human researchers and whether they provide useful insights into the information.
“AI can help researchers see patterns across large amounts of information, but the real question is what those patterns tell us about people’s experiences,” said Chau Tong, an assistant professor in the Missouri School of Journalism with a joint appointment in the MU Institute for Data Science and Informatics. “By bringing together different areas of expertise, we can better evaluate how these tools support research and where human judgment remains essential.”
Putting AI’s analysis to the test
The team tested the evaluation framework, called the Alignment and Usefulness Rating Audit (AURA), by analyzing 32 interviews with nursing faculty, practitioners and students about their experiences with generative AI.
Researchers first identified key themes in the interviews, then compared their findings with topics generated by a computer model. AI identified many of the same broad themes as human researchers, but it often missed important distinctions in meaning and context.
Eleven of the 12 themes identified by human researchers closely matched at least one topic generated by the computer model. However, a match did not necessarily mean the model captured the full meaning of a theme or organized the information in the most useful way.
The model sometimes combined several distinct themes into one broad topic or divided a single theme into several smaller ones.
“When people talk in interviews, they don’t usually stay on one topic from start to finish,” said John Robert Bautista, an assistant professor in Mizzou’s Sinclair School of Nursing with a joint appointment in the MU Institute for Data Science and Informatics. “AI can help organize those conversations so researchers can more easily see what people are saying.”
Where AI helps and where it falls short
While AI helped researchers organize large amounts of information and identify patterns, the team found less evidence that it consistently generated meaningful new insights beyond what human analysts identified.
“Our findings suggest the technology can help researchers manage large collections of interviews, focus groups and survey responses,” said Temitope Eyinla, a doctoral researcher in the MU Institute for Data Science and Informatics. “But human expertise remains central to interpreting the results.”
The evaluation could support research into issues such as patient experiences, educational challenges and workplace concerns, where understanding people’s perspectives is essential. However, the study examined a relatively small set of interviews, so additional testing is needed to determine how well AURA works across different subject areas and types of research.
The European Laboratory for Artificial Intelligence (ELAI) has won the call by the European Commission to pilot a virtual institute of world-class excellence under the Resource for AI Science in Europe (RAISE). The announcement was made today in Brussels at the Apply AI Summit in the European Commission and Martin Brynskov, ELAI Coordinator from the University of Copenhagen.
ELAI brings together a consortium of 17 sites, 33 partners, and 21 supporters in 17 countries, including Japan and Canada, on 3 continents, to establish a distributed virtual institute, built on excellent research institutions creating the structure for leading researchers to tackle, at the European level, Frontier AI research challenges, including in efficiency, safety and security. With €17M seed funding over five years, it will amplify a strategic research agenda that pools and provisions resources for scientific breakthroughs to ensure European leadership in AI research and innovation to deliver on societal challenges.
'ELAI unites leading AI research institutions combining world-class fundamental AI research with deep expertise in machine learning for the sciences. Building on almost a decade of collaboration through the European Laboratory for Learning and Intelligent Systems (ELLIS), we have proven ways to organise joint research and attract internationally leading talent at a scale no single site or national effort could achieve alone,' says Serge Belongie, Professor at University of Copenhagen, Director of the Pioneer Centre for Artificial Intelligence, and Board President of ELLIS.
The announcement was made at the Apply AI Summit in Brussels, which marks one year since the adoption of the Apply AI Strategy and concludes the first European AI Innovation Month. The Summit showcases how trustworthy, human-centric and innovative AI can boost the competitiveness of sectors such as healthcare, mobility and the public sector, with European use cases already running on homegrown AI models and systems.
RAISE is a pilot launched by the European Commission to deliver on the AI in Science strategy, which supports and incentivises the development and use of AI by the European scientific community. It pools strategic resources (funding, compute, data and talent) and operates along two pillars: Science for AI, which supports basic research to advance core AI capabilities, in particular safe and secure frontier AI; and AI in Science, which promotes the use of AI for progress in different scientific disciplines. ELAI is funded under the Science for AI pillar.
ELAI’s activities will begin on 1 January 2027.
About ELAI The European Laboratory for Artificial Intelligence (ELAI) establishes a distributed, excellence-driven virtual institute that organises leading researchers around a shared frontier research agenda to advance science for AI, establishing new foundations for efficient, reliable and interactive learning systems. It is funded by the European Commission and led by the University of Copenhagen.
AI system developed to recommend content for improving emotional health
Imagen de la aplicación EMOTINET, que recomienda contenidos adaptados para mejorar su salud emocional. Aparece un chatbot que pregunta: MarÃa, selecciona los temas que te interesarÃa trabajar en Emotinet
Researchers at the Universidad Carlos III de Madrid (UC3M), in collaboration with the company Psiconnea, have developed an artificial intelligence (AI)-based system capable of detecting emotional patterns in individuals and recommending tailored content to improve their emotional well-being. The platform, available via the EMOTINET app, enables users to perform self-assessments to spot early signs of psychological distress and—through exercises, visualizations, and professionally guided series—overcome mental blocks, reduce anxiety, and gain emotional clarity.
To achieve this, the app developed by Psiconnea (a Spanish company of psychotechnologists specializing in emotional health) processes various types of data, including the user's psychological profile, their interaction with the platform, and their self-reported needs. The AI system designed by UC3M integrates all this information to generate personalized multimedia recommendations. Specifically, these consist of brief explanatory videos—acting as "micro-learning modules"—that help users better address their specific challenges and have been pre-validated by Psiconnea.
EMOTINET also features emotional self-assessments, psychologist-guided exercises, and an anonymous support community. According to Psiconnea CEO Rosa Becerril, more than 80,000 people have used the platform, and 80% of users reduce their levels of stress, anxiety, or depression.
Seven research-based practices offer steps for evaluating and introducing AI tools with a focus on trust, human oversight, job design and workers’ well-being
WASHINGTON — As organizations introduce artificial intelligence tools to their workforce, people may be left to navigate unclear policies, conflicting expectations and new questions about trust and acceptable AI use in the workplace.
A new guide released by the American Psychological Association outlines what organizations should consider when introducing AI into people’s work.
Created for organizations seeking to adopt AI responsibly while also supporting employees and improving how work is designed, Deploying AI Into Workplaces: What Leaders Need to Know offers seven best practices based on established psychological research and emerging evidence about AI. These range from matching safeguards to the stakes of potential errors to ensuring that AI is introduced through a fair process.
“The rapidly growing use of AI at work is creating new behaviors and social judgments while raising questions about trust, all of which are fundamentally questions of psychology. Did you use AI to write your emails today? Am I supposed to think more or less of you if you did?” said APA CEO Arthur C. Evans Jr., Ph.D. “Psychological science can help organizations understand not just what AI tools can do, but how their use affects trust, decision-making and employee well-being. APA’s new guide gives leaders a practical framework for introducing AI thoughtfully, with people at the center of the process.”
The share of workers reporting pressure from their employers to use AI at work is rising. Yet, according to APA’s 2026 Work in America Survey released earlier this year, less than 40% of workers for whom this question applied reported knowing whether their employer had an official policy about acceptable AI use.
To address the adoption-policy mismatch, APA’s new guide offers a practical checklist with seven ways to help organizations introduce, use and evaluate AI tools responsibly. It also offers examples of what effective AI tool implementation can look like, as well as identifies the red flags to watch out for. Rather than providing universal instructions or telling organizations whether or not to use AI tools, the guide is designed to help organizations that choose to introduce these tools do so thoughtfully.
Developed with input from an advisory panel of experts in human-technology interaction and clinical, social, legal, cognitive and industrial-organizational psychology, the guide’s suggested practices include:
Matching safeguards to the stakes of potential errors and thinking through consequences and scenarios that may lie ahead;
Ensuring that AI is introduced through a fair process, including consulting with those who will be using the new system and those who will be affected by the outputs of that system; and
Protecting opportunities for people to develop deep expertise, share knowledge and learn from experience.
“Purchasing an AI tool and telling people to use it without further direction is not a sound strategy,” said Dennis P. Stolle, J.D., Ph.D., executive lead psychologist for applied psychology at APA. “Organizations need to invest in the people who will work with the technology, not just the technology itself. That means giving workers the information they need to judge where AI tools are reliable and where they are not, making sure human oversight is meaningful and rethinking how jobs are designed.”