The central problem the authors identify is one of traceability. When a patient receives bad advice from a physician, there is a medical record, a licensing board, an insurance claim, and often a malpractice pathway that together leave a documented trail. When a patient receives bad advice from a chatbot, the conversation typically remains locked inside the AI company’s platform, the user has no formal mechanism for reporting the problem, and outside researchers have no way to audit what was said or to estimate how often things go wrong. Kaicheng Yang, an assistant professor in the School of Computing at Binghamton University’s Thomas J. Watson College of Engineering and Applied Science, put the situation bluntly: people are getting medical advice from chatbots, but nobody is looking at it, and if there is an error or an issue there is simply no way for anyone to know.

The paper illustrates the danger with a striking real-world case. A 60-year-old man asked ChatGPT how to reduce chloride in his diet, and based on the guidance he received, he replaced table salt with sodium bromide, a chemical cousin that was once used industrially but is toxic to the human nervous system. He was ultimately hospitalized with bromide toxicity after developing hallucinations and paranoia, symptoms that can mimic psychiatric illness and that historically required careful detective work to attribute to bromide exposure. The critical failure in this case, the authors argue, was not only the advice itself but the information vacuum that followed. Because the man could not retrieve the conversation he had with the chatbot, his physicians had no way of knowing what he had been told, making it nearly impossible to reconstruct the pathway that led to his hospitalization.

This information vacuum, the researchers contend, actively prevents oversight rather than merely failing to invite it. In their framing, the harms of AI-generated health advice are not only possible but structurally hidden from the very clinicians, researchers, and regulators who could otherwise detect and correct them. The distinction matters because it reframes the issue from one of occasional bad outputs to one of systemic accountability. A single erroneous answer from a chatbot is a quality-control problem; an ecosystem in which erroneous answers leave no verifiable trace is a governance problem. Without visibility into what chatbots actually tell people about their health, there can be no meaningful epidemiology of AI-related harm, no feedback loop for improving models, and no evidence base for regulation.

The paper also emphasizes that chatbots can be wrong in characteristic ways that make the problem worse. Large language models can generate erroneous answers with unwavering confidence, a property that makes mistakes harder for lay users to spot. They can oversimplify complex medical questions that require individualized clinical judgment. They can draw on outdated information or reproduce false claims that circulated widely in their training data. Each of these failure modes has been documented in evaluations of AI systems on medical question benchmarks, and each becomes more consequential when users treat the output as authoritative guidance rather than as a starting point for a conversation with a qualified professional.

Perhaps most provocatively, the authors argue that these risks no longer require users to seek out health advice at all. AI-generated summaries now appear at the top of search engine results pages, and AI assistants are being integrated directly into social media platforms such as X and Meta’s services. This means that a person scrolling through a feed can be exposed to health claims they never asked for, generated by systems with the same failure modes as standalone chatbots but delivered in a context where users are even less likely to scrutinize them. Yang noted that he is sometimes not even looking for health information, yet it simply shows up while browsing social media feeds, and he and his colleagues believe this could produce undesirable outcomes, particularly when medical misinformation or state actors attempting to manipulate online discussion enter the picture.

Whether a person actively solicits medical guidance from a chatbot or encounters AI-generated health content incidentally, the paper argues that the resulting harm rarely leaves traces that clinicians, regulators, or researchers can verify. This asymmetry between the scale of AI health information delivery and the near-total absence of monitoring infrastructure is what gives the paper its urgency. Traditional public health surveillance depends on detectable signals: emergency room visits, laboratory findings, adverse event reports, patterns in insurance data. The bromide toxicity case was detected only because the patient’s symptoms were severe enough to require hospitalization and because clinicians eventually identified the chemical cause. Countless subtler harms, such as delayed care, inappropriate self-treatment, or misplaced reassurance, would generate no such signal.

To close this gap, the authors propose a package of transparency and accountability measures aimed at different points in the ecosystem. AI companies, they argue, should give users access to their own health-related conversations so that patients can share them with clinicians, allowing physicians to investigate the pathway that led to a harmful event. Those companies should also develop disclosure systems through which health guidance can be flagged, reported, and formally investigated, creating the kind of adverse event reporting infrastructure that exists in pharmacovigilance and medical device regulation. Social media companies, in turn, should take greater responsibility by clearly labeling AI-generated health content and withholding such content until it has been vetted by medical governing bodies. Search engines should restrict the information used in AI-generated summaries to vetted sources. Policymakers, the authors suggest, should consider extending physician malpractice liability frameworks to AI chatbot companies, aligning the legal exposure of these systems with the medical role they have effectively assumed in many people’s lives.

The authors are skeptical that voluntary corporate action will suffice. Yang stated that this is not something society should rely on the companies to do, because their incentive is always to make more money, and building a robust transparency system goes against that incentive. Instead, he argued for some kind of third-party monitoring system, an independent evaluation mechanism that can examine how AI systems handle health queries without depending on the platforms themselves to disclose their own failures. This position echoes long-standing arguments in drug and device regulation, where independent oversight emerged precisely because manufacturers could not be trusted to police themselves when safety and profit conflicted.

The paper arrives at a moment when the volume of AI-mediated health information is growing faster than any mechanism for studying it. As chatbots become embedded in search engines, operating systems, and social networks, the boundary between seeking medical advice and simply consuming content continues to dissolve, and with it the last remaining points where harm could be observed and documented. The authors’ argument is ultimately a call to treat AI-generated health information as a public health exposure worthy of surveillance, reporting requirements, and independent audit. Without such measures, the damage caused by bad chatbot advice will remain what the paper’s title suggests it already is: invisible, not because it does not occur, but because no one has built the instruments to see it.

Subject of Research: The invisible risks and lack of oversight of AI-generated health information from chatbots

Article Title: When AI chatbots give bad advice, no one can see the damage

Article References: When AI chatbots give bad advice, no one can see the damage. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: AI chatbots, health misinformation, Nature Health, medical advice, transparency, regulation, patient safety, large language models, social media, accountability, public health, Binghamton University

The trial was ambitious in scale and design. Between April and June 2025, the research team recruited 2,400 adults with no formal medical education across 24 healthcare systems in China. Participants were randomly assigned in a one-to-one ratio to either the LungDiag chatbot or a browser-based search control, with dedicated AI and generative AI-assisted search functions deliberately disabled in the control arm to ensure a fair comparison. The study was single-blind, prospective, non-interventional and multicentre, and it was preregistered, a detail that matters enormously in a field where retrospective evaluations of large language models have often been criticized for flexibility in how results are reported. Each participant worked through simulated clinical vignettes describing respiratory conditions, and the primary endpoint was item-level accuracy across four domains: identifying risk factors and aetiology, reaching a preliminary diagnosis, and deciding on the appropriate triage level.

The statistical machinery behind the analysis was correspondingly rigorous. Rather than simply pooling correct and incorrect answers, the investigators used item-level generalized linear mixed-effects models with random intercepts for both participants and items, an approach that accounts for the fact that some people are naturally better at these tasks than others and that some vignettes are inherently harder. In the primary compliant-case analysis, 2,176 questionnaires were included, with 1,088 in each group. The result was clear: LungDiag achieved a model-based overall accuracy of 70.0 percent, with a 95 percent confidence interval of 65.2 to 74.8 percent, compared with 55.4 percent for the web search control, whose interval ran from 49.7 to 61.1 percent. That corresponds to a risk difference of 14.6 percentage points, a risk ratio of 1.26, and a P value below 0.001.

What is perhaps most striking is where the advantage came from. The largest gap between the two tools appeared on preliminary diagnosis, where the chatbot-assisted group reached 81.2 percent accuracy against just 54.5 percent for the search group, a difference of nearly 27 percentage points. In other words, when the task was to figure out what condition a described patient might have, the constrained AI assistant was dramatically more effective at guiding laypeople to the right answer. Gains on triage, the task of deciding how urgently care is needed, were real but more modest: 53.4 percent versus 47.1 percent. This asymmetry is scientifically interesting because it suggests that a curated respiratory knowledge layer layered on top of a general-purpose language model can sharpen diagnostic reasoning more readily than it can resolve the genuinely difficult judgment calls about urgency that even clinicians sometimes disagree on.

That triage limitation deserves close attention. For binary high-urgency triage, deciding whether a case requires immediate emergency attention, LungDiag showed high sensitivity at 85.6 percent but only moderate specificity at 62.1 percent. The control arm performed worse on both counts, at 80.9 percent and 56.0 percent respectively. High sensitivity means the chatbot rarely missed a true emergency, which is the safer failure mode, but the moderate specificity implies it flagged many non-emergencies as urgent. In a real-world deployment, that trade-off could funnel worried-well users toward emergency departments that are already strained, a consequence the authors themselves acknowledge the study cannot fully evaluate. The trial was a simulated-case comparison, and the researchers are explicit that the findings do not establish effects on actual care-seeking behavior, diagnostic delay, or clinical outcomes.

There is also a cost to the accuracy, and it is measured in seconds. Median task completion time was 527 seconds for the LungDiag group versus 461 seconds for the control, roughly a minute longer per case. In a triage context, where a user may be anxious or where minutes could matter, that overhead is not trivial. It likely reflects the conversational nature of the tool, which asks clarifying questions and walks users through structured reasoning rather than dumping a list of links. Whether users would tolerate that friction in real life, and whether the extra minute buys enough accuracy to be worth it, are questions the trial design cannot answer but that any deployment would need to weigh.

The architecture of LungDiag is what separates it from simply asking a chatbot for medical advice. The system is task-constrained, meaning it is deliberately limited to respiratory assessment rather than open-ended conversation, and it incorporates a respiratory knowledge layer derived from the team’s earlier LungDiag work on electronic health records across multiple centers. This retrieval-augmented approach grounds the model’s outputs in curated clinical resources rather than relying solely on patterns learned during pretraining, which is one plausible explanation for why laypeople guided by the tool made fewer diagnostic errors than those sifting through search results on their own. The exact internal prompt text, retrieval-engine settings, and knowledge layer resources were not publicly released because they contain proprietary implementation details, though the functional specification and architecture are described in the supplementary materials, and de-identified participant-level datasets and analysis code are openly available on GitHub and Zenodo.

The context of respiratory disease in China gives the study particular urgency. Chronic respiratory conditions impose an enormous and growing burden globally, and lower respiratory infections remain among the leading causes of death worldwide according to the Global Burden of Disease project. In a country where WeChat functions as a near-universal communication layer, embedding a competent triage assistant directly into an app that more than a billion people already open every day could reach populations that formal healthcare access struggles to serve, particularly in rural regions. The trial’s 24-site footprint across diverse Chinese healthcare systems, and subgroup analyses by sex, age, education and region, suggest the team was attentive to exactly this question of generalizability.

The study also sits within a rapidly accumulating evidence base on human-AI interaction in medicine. Recent randomized trials have shown that GPT-4 assistance can improve physician performance on patient care tasks, while other work has documented both the promise and the pitfalls of large language models in clinical decision-making, including failures on tasks requiring integration of longitudinal context. What distinguishes the new study is its focus on laypeople rather than clinicians, and its use of a nationwide messaging platform rather than a standalone application. Most people who seek health information online will never consult a medical AI through a dedicated clinical portal; they will use whatever is already on their phone. Testing the technology in that realistic environment, against the realistic alternative of web search, is what makes the 14.6 percentage-point improvement meaningful.

Cautions remain, and the authors do not shy away from them. Simulated vignettes are not sick patients; they compress clinical reality into text and cannot capture the physical examination findings, hesitation, and ambiguity of real presentations. Accuracy on vignettes does not guarantee that a chatbot will reduce dangerous delays in seeking care, and questions of medical liability, regulatory oversight, and the black-box nature of large language models loom over any deployment. Still, the trial sets a methodological benchmark: a preregistered, randomized, adequately powered, transparently analyzed comparison with public data and code. As health systems worldwide grapple with surging demand and clinician shortages, the message from this study is cautiously optimistic. A carefully constrained AI assistant, built on a general-purpose model but grounded in domain knowledge and delivered through the messaging apps people already use, can help non-experts reason about respiratory illness substantially better than the search bar they would otherwise turn to. The next test is whether that advantage survives contact with the messy reality of actual patients, actual anxiety, and actual emergency rooms.

Subject of Research: Accuracy of a GPT-4o-based respiratory triage chatbot versus web search for layperson diagnosis and triage

Article Title: Accuracy of a respiratory assessment chatbot in a nationwide messaging system for layperson diagnosis and triage: a randomized preregistered study

Article References: Liang, J., Wang, Y., Ni, J., Cai, Y., Chen, K., Wang, J., Shi, Y., Chen, B., Dong, C., Guo, T., Huang, J., Huang, Z., Guo, Z., Li, J., Liu, B., Pu, Z., Qi, S., Sun, R., Wang, R., … He, J. (2026). Accuracy of a respiratory assessment chatbot in a nationwide messaging system for layperson diagnosis and triage: a randomized preregistered study. Nature Health. https://doi.org/10.1038/s44360-026-00189-9

Image Credits: AI Generated

DOI: 10.1038/s44360-026-00189-9

Keywords: artificial intelligence, chatbot, GPT-4o, respiratory disease, triage, diagnosis, randomized controlled trial, WeChat, large language models, digital health, layperson health literacy, Nature Health

The research was led by teams at the Center for Climate, Health, and the Global Environment at Harvard T.H. Chan School of Public Health and the Department of Epidemiology at Boston University School of Public Health, in collaboration with the health data company Truveta. Their approach represents a proof of concept for what the authors describe as near-real time surveillance of climate-related health harms. Rather than waiting for administrative claims to be processed and released, the team drew directly on clinical records generated at the moment patients walked into emergency departments, transforming routine medical documentation into actionable intelligence for an ongoing climate event.

The scale of the analysis was substantial. The researchers examined 8.1 million emergency department visits recorded during the summers of 2023 through 2025, drawing on de-identified electronic health records contributed by Truveta member health systems spanning 19 Eastern U.S. states. They linked daily maximum temperatures to state-level counts of emergency department visits for all causes, heat-related illnesses, kidney diseases, and psychiatric conditions. From these connections, they estimated how much visit rates rose on hotter days and calculated how many visits during the June and August 2025 heat waves could be attributed to extreme temperatures.

The findings quantify, with unusual immediacy, just how heavily heat weighs on emergency care. During the June 2025 heat wave, approximately 9.2 percent of all-cause emergency department visits were attributable to heat, corresponding to roughly 6,986 additional visits across the 19 states studied. The August 2025 heat wave produced a similar burden, with about 8.3 percent of all-cause visits attributable to heat, or approximately 6,351 additional visits. Taken together, the two events were responsible for more than 13,000 excess emergency department visits, a surge that the authors argue demands real-time visibility and rapid response rather than retrospective accounting.

Perhaps the most striking pattern emerged in kidney disease. Nearly 30 percent of kidney disease-related emergency department visits during heat wave periods were attributable to heat, a proportion that underscores how profoundly high temperatures stress the body’s fluid and thermoregulatory systems. Across the entire warm season from April through August in 2023 to 2025, warmer temperatures were associated with a 6.3 percent increase in all-cause emergency department utilization, a 19.5 percent increase in visits for kidney diseases, and an 8.8 percent increase in visits for psychiatric conditions. These associations held across three consecutive summers, suggesting a consistent and predictable relationship between rising temperatures and acute health care demand.

One finding in particular carries weight for how health agencies evaluate their own programs. Intuition might suggest that heat-related emergency visits would ease as summer progresses, as residents acclimatize and heat advisories become routine. The data showed only a modest decline between the June and August heat waves. According to Amruta Nori-Sarma, the study’s lead author and deputy director of Harvard Chan C-CHANGE, this has important implications for real-time evaluation of locally implemented adaptation programs. If adaptation measures such as cooling centers or outreach campaigns are working, their effect should be visible in emergency department patterns within the same season, and if they are not working, agencies can adjust before the next event rather than the next year.

The practical applications the researchers envision extend across the preparedness landscape. Near-real-time emergency department data can validate whether heat action plans and early warning systems are actually reducing health burdens, for example by confirming whether opening cooling centers coincides with fewer heat-related visits. Rapid analyses can help target urban greening and cool infrastructure investments toward the neighborhoods and populations showing the strongest spikes during heat. Electronic health record surveillance can also highlight which groups, such as outdoor workers, older adults, or people with chronic kidney disease, need tailored public health messaging and worker protections as upcoming heat events approach. By comparing emergency department patterns before and after new measures are introduced, health agencies can see within a single summer whether programs are protecting patients.

Underlying the study is a technological shift in how clinical data can be mobilized for public health. The records used came from Truveta, a platform that aggregates de-identified electronic health records from member health systems, and were available within one week of each emergency department encounter. That turnaround, compared with the months-long lag of insurance claims, changes the fundamental character of heat surveillance: it moves from forensic reconstruction to situational awareness. As Brianna Cartwright, director and lead of Truveta Research and a co-author on the paper, noted, timely real-world data allow researchers to rapidly assess the health impacts of extreme heat across diverse populations and geographies, providing a clearer picture of risk and the potential to transform how climate-related health emergencies are monitored, understood, and answered.

The policy implications outlined by the authors are direct. They call for investment in data systems that permit timely, secure use of electronic health records for public health situational awareness, and for the integration of near-real-time emergency department surveillance into heat action plans, emergency preparedness frameworks, and broader climate resilience efforts. They also urge focused attention on the populations the data identify as most at risk, especially older adults and people with kidney disease. As Mary Willis, senior author and assistant professor of epidemiology and environmental health at Boston University School of Public Health, emphasized, the surge of more than 13,000 excess visits from just two heat waves underscores the importance of investing in timely population health surveillance and corresponding intervention strategies in an era of increasing extreme heat.

As climate change lengthens and intensifies heat seasons across the United States and beyond, the gap between harm and response has become one of the most consequential variables in public health. This study suggests that the gap can be closed not by new medical treatments but by faster information: a shift in how existing clinical data are collected, linked, and interpreted. If the approach demonstrated across 19 states can be scaled and sustained, the summers ahead may be the first in which health systems can watch a heat wave’s impact unfold in real time, redirect resources toward the patients and communities being hit hardest, and learn within weeks rather than years which protective measures actually save lives.

Subject of Research: Near-real-time surveillance of heat wave health impacts using electronic health records

Article Title: Rapid response to heat waves: Using electronic health records to protect patients in near-real time

Article References: Rapid response to heat waves: Using electronic health records to protect patients in near-real time. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: heat waves, electronic health records, emergency department visits, public health surveillance, extreme heat, kidney disease, climate change, Nature Health, Harvard Chan C-CHANGE, Boston University School of Public Health, Truveta, heat action plans

The researchers set out to map where, in the machinery of modern healthcare delivery, AI-related harms actually occur. Rather than treating AI as a single black box that a doctor either accepts or rejects, they categorized the cases according to the specific clinical tasks physicians perform: diagnosing conditions, ordering and interpreting tests, planning treatment, triaging patients, documenting and communicating information, and coordinating care across institutions. The distribution was revealing. Harms clustered not only around diagnostic tools — the category that dominates public imagination and much of the regulatory conversation — but also around administrative and coverage systems, including AI used by insurers to deny or curate care, and screening tools deployed by state agencies to flag families for investigation.

That breadth matters because it exposes a structural blind spot in both technology design and legal doctrine. Much of the existing scholarship on AI liability in medicine has centered on a deceptively simple question: if an AI tool outperforms a doctor, does reliance on it constitute negligence, or does deviation from it? Prior work, including widely cited analyses in JAMA and law reviews, has debated whether physicians could face liability for ignoring a competent algorithm or for deferring to a flawed one. Mansi and Riedl argue that this physician-centered frame is too narrow to describe what actually happens when patients are harmed. An insurance company’s algorithm may terminate home-care coverage; a state screening tool may incorrectly flag a parent for child-welfare investigation; a triage system may misdirect emergency care. In each scenario, the patient’s injury emerges from decisions distributed across an organization, with the physician often having little visibility into — or control over — how the AI was configured, trained, deployed or applied.

The legal cases the researchers analyzed come from a meticulous triangulation of three public databases: the Database of AI Litigation maintained at George Washington University Law School, the Health Litigation Tracker, and the AI, Algorithmic and Automation Incidents and Controversies repository, a community-maintained archive of documented harms. As of February 2025, all three resources are freely accessible online, and the researchers referenced individual cases through unique identifiers, centering documented harms rather than the specific systems involved. The case-study approach allowed the authors to reconstruct, case by case, how harm materialized, which stakeholders were positioned between the patient and the algorithm, and what barriers the patient encountered in seeking recourse.

Those barriers, the study finds, are both legal and technological, and they reinforce one another. On the legal side, liability structures remain tethered to a model of medical malpractice that presumes a human decision-maker with enough information and discretion to have acted differently. When an algorithm’s output is embedded in a payer’s coverage workflow or a state agency’s screening pipeline, a patient’s lawyer must identify which entity’s conduct was negligent, prove causation, and often overcome trade-secret protections that keep the model’s logic opaque. On the technological side, the tools themselves are rarely designed for contestability. Explanations, where they exist, are built for clinicians — and even for clinicians, a substantial body of human-computer interaction research shows that explainability features are frequently underused, misunderstood or shaped more by regulatory checkbox than by genuine decision support. They are almost never built for the people who most need them when things go wrong: patients and their advocates.

Here the study makes its most provocative move. Mansi and Riedl argue that lawyers working for patients should be recognized as legitimate users of AI systems — and, in a sense, as intermediaries whose needs deserve to shape the design of medical AI from the ground up. If an explainability interface cannot help a patient’s legal team understand which stakeholder made which consequential decision, what data the model relied on, and where in the pipeline the error propagated, then the tool has effectively insulated everyone but the patient from scrutiny. The authors connect this to a growing research program on legally-informed explainable AI, which aims to produce explanations aligned with the evidentiary demands of legal proceedings rather than with the aesthetic of transparency that currently dominates industry practice.

The researchers also draw a pointed lesson from documented incidents involving coverage algorithms. Reporting on how Medicare Advantage plans have used AI tools to cut off care for seniors, alongside legal scholarship on regulating healthcare coverage algorithms, illustrates the pattern the study identifies: harms arise at the intersection of commercial incentives, algorithmic automation and fragmented oversight. A patient denied post-acute care by an algorithmic coverage tool may appeal through administrative channels, sue, or simply suffer silently. Each path demands different evidence, different expertise and different institutional access — resources that are unevenly distributed among the very populations most likely to be harmed.

What can be done? The authors propose reform on two fronts. First, liability structures should evolve to reflect the distributed reality of AI-mediated care. They point to legal scholarship exploring theories such as the common enterprise theory of liability, which could allow responsibility to be shared across the ecosystem of actors — developers, payers, facilities, regulators — that jointly shape how a tool affects patients. Distributed governance frameworks for medical AI, and regulatory thinking that treats software as a medical device within a whole system rather than as a standalone artifact, offer templates for how accountability could be allocated more realistically. Second, AI tools for healthcare should be designed for contestability by default: systems should log which stakeholders shaped their outputs, generate explanations usable in legal contexts, and support advocates — including patients’ lawyers — in reconstructing what happened when harm occurs. Contestable AI, an emerging design philosophy in human-computer interaction, offers concrete principles: make decisions reversible, make the grounds of decisions accessible, and make the pathway to human review visible to affected people.

The stakes of getting this wrong extend beyond individual lawsuits. If patients cannot seek recourse, there is no corrective feedback signal pushing developers and deployers to fix flawed systems; courts become the only place where the accumulated evidence of algorithmic failure can surface, and only for those few patients wealthy or persistent enough to get there. Conversely, if accountability is designed into the system — technologically through auditable, contestable tools, and legally through liability frameworks that track the actual distribution of decision-making power — then recourse becomes possible without requiring every patient to become a litigant. The study’s authors, whose work was supported in part by the National Science Foundation, frame this as a shift from physician-centered to patient-centered accountability, a reorientation that demands new collaboration between lawyers, technologists, clinicians and regulators.

For a healthcare system rapidly absorbing AI into triage, diagnosis, documentation and coverage decisions, the message of the litigation record is uncomfortable but clear: the technology is arriving faster than the accountability structures needed to govern it, and the people bearing the consequences are the ones the current framework serves least. Mansi and Riedl’s analysis of 31 cases turns courtroom silence into a design specification — a map of where AI harms happen in healthcare delivery, and a blueprint for tools and laws that could finally let patients fight back.

Subject of Research: Analysis of US litigation involving medical AI tools and patient-centered accountability for AI-related harms in healthcare delivery.

Article Title: Implications of current litigation on the design of AI tools for healthcare delivery and related legal frameworks

Article References: Mansi, G., & Riedl, M. (2026). Implications of current litigation on the design of AI tools for healthcare delivery and related legal frameworks. Nature Health. https://doi.org/10.1038/s44360-026-00186-y

Image Credits: AI Generated

DOI: 10.1038/s44360-026-00186-y

Keywords: medical AI, healthcare litigation, AI liability, patient recourse, explainable AI, contestable AI, healthcare delivery, algorithmic harms, coverage algorithms, legal frameworks, patient-centered accountability, Nature Health

The study examined Burundi, Ethiopia, Malawi, Mozambique and Namibia, five countries that differ enormously in wealth, food security and education system maturity, yet all of which operate national or large-scale school feeding programmes. The researchers combined nationally representative household surveys with administrative education data, comparing children who received school meals with children who did not, both at the national level and across subnational regions, and separately for boys and girls. Their central finding is striking: beneficiaries of school feeding would gain, on average, 1.3 additional years of schooling compared with non-beneficiaries, with country-level estimates ranging from a low of 0.6 years to a high of 2.3 years.

The logic behind these gains is grounded in decades of nutrition and education research. Chronic food insecurity pushes families toward painful trade-offs, and one of the most common responses is to withdraw children from school, either because the family cannot afford the indirect costs of attendance or because a hungry child cannot concentrate, walk long distances to class or sustain the physical demands of the school day. School meals act as a form of social protection, transferring resources directly to households while strengthening the incentive to keep children enrolled. When a nutritious meal arrives reliably at school, the marginal cost of sending a child falls, and the marginal benefit rises, making dropout less likely and grade progression more likely.

Translating additional schooling into money is where the study becomes an economic argument as much as an educational one. The researchers estimated the returns to education using adapted Mincer equations, the workhorse framework of labour economics that relates log earnings to years of schooling and experience, supplemented where appropriate by published estimates from the global literature. Across the five countries, each additional year of education was associated, on average, with 13.3 per cent higher earnings, with estimates ranging from 9.5 per cent to 19.5 per cent. Combining the schooling gains with these returns, the team projected that school feeding beneficiaries would see lifetime earnings approximately 25.6 per cent higher than comparable children who did not receive meals, an enormous cumulative effect originating from what is often regarded as a modest, routine intervention.

The disaggregated analysis reveals where those benefits concentrate. Because the researchers modelled impacts both nationally and subnationally, they could identify the regions within each country where school feeding is likely to deliver the largest educational payoffs. Unsurprisingly, the greatest gains appear in disadvantaged areas, where baseline dropout rates are high, food insecurity is severe and alternative social protection is thin. In such settings, the meal programme effectively compensates for broader gaps in the social safety net. The analysis by sex also allows the programme’s role in closing gender gaps in schooling to be examined, an important consideration given that girls in low-income settings are often the first to be withdrawn from school when household resources tighten, and given the well-documented downstream benefits of girls’ education for fertility, child mortality and intergenerational health.

The methodological machinery behind these estimates is worth unpacking, because the authors are candid about its limits. In Ethiopia and Malawi, where rich household survey data were available, including the Ethiopian Socioeconomic Survey and multiple rounds of the Malawi Integrated Household Survey, the team estimated transitions between schooling states directly from observed data. For Burundi, Mozambique and Namibia, they constructed a documented transition-matrix framework, using education management information system data and national statistics to model how beneficiary and non-beneficiary cohorts progress through the school system. The schooling gains were then propagated through Mincerian or literature-based returns to produce projected earnings differentials. The analytic code, along with processing and estimation pipelines for the countries with public data, has been released openly on GitHub, an unusually transparent step that allows other researchers to reproduce, stress-test and extend the modelling.

That transparency matters because the authors themselves label their headline numbers as observational plausible upper bounds rather than causal guarantees. Beneficiaries and non-beneficiaries are not randomly assigned; school feeding programmes are typically targeted, deliberately or otherwise, toward needier regions, which complicates naive comparisons. The team ran extensive sensitivity analyses, visualised in their published heat maps, to explore how alternative assumptions about returns to education, transition probabilities and programme coverage affect the results, and the qualitative conclusion, that school feeding meaningfully increases retention and lifetime earnings, is robust across specifications. Even so, they explicitly call for new longitudinal quasi-experimental studies, ideally exploiting staggered programme rollouts or sharp eligibility thresholds, to pin down causal estimates with the precision that policy commitments of this scale deserve.

Even interpreted as upper bounds, the findings arrive at a consequential moment. School feeding has expanded dramatically worldwide, reaching hundreds of millions of children, and is increasingly championed not merely as a nutrition intervention but as a cross-sectoral investment spanning health, education and social protection. Previous systematic reviews and meta-analyses have found positive but heterogeneous effects of school meals on enrolment, attendance and learning, and value-for-money studies have been completed or are underway in each of the five study countries under the Research Consortium for School Health and Nutrition, funded through an original grant from the Norwegian Agency for Development Cooperation with follow-on support from the World Food Programme and other partners. The new Nature Health analysis adds a lifetime-earnings dimension to that evidence base, converting years of schooling into the currency that finance ministries ultimately weigh.

The economic framing also reframes the cost debate. School feeding programmes require sustained procurement of food, logistics and administration, and critics have long questioned whether the resources might deliver more if spent on textbooks, teacher training or direct cash transfers. But the earnings projections here suggest that the benefits of keeping children in school compound over entire working lives, and that they accrue disproportionately to the poorest regions, where the marginal social value of a retained pupil is highest. There are broader dividends as well, beyond the individual wage premium: parental education is strongly linked to child survival, and education shifts fertility patterns and population health in ways that markets do not fully capture. A meal that keeps a girl in school for two extra years may influence the health and schooling of the next generation.

For the five countries studied, and for the many others watching, the message is that school meals should be evaluated with the same seriousness as any major public investment, using data on who benefits, where the gains are largest and what the long-run returns look like. The Harvard-led team’s open data and code make that kind of evaluation more feasible than ever. What remains is the harder scientific task they themselves identify: following real cohorts of children over time, with designs that isolate the causal effect of the meal from the poverty that motivates it. Until then, the numbers stand as a provocative estimate that one of the oldest interventions in the development toolkit, feeding hungry children at school, may quietly be one of the highest-return investments a government can make.

Subject of Research: Estimated effects of school feeding programmes on educational attainment and lifetime economic returns in Burundi, Ethiopia, Malawi, Mozambique and Namibia

Article Title: Impact of school feeding programmes on educational attainment and economic returns in five African countries

Article References: Forzy, T., Ramponi, F., Iversen, I., Durizzo, K., Kim, S., Gautam, P., Avallone, S., Memirie, S. T., Habtemichael, M., Getnet, F., Wogasso, Y., Peterson, H., Masamba, K., Saka, A., Nkengurutse, J., Ndayitwayeko, W.-M., Ntunzwenimana, M., Bigirimana, L., Mbengue, M., … Verguet, S. (2026). Impact of school feeding programmes on educational attainment and economic returns in five African countries. Nature Health. https://doi.org/10.1038/s44360-026-00176-0

Image Credits: AI Generated

DOI: 10.1038/s44360-026-00176-0

Keywords: school feeding, educational attainment, lifetime earnings, sub-Saharan Africa, returns to education, dropout, social protection, Mincer equation, food security, Nature Health, Harvard Chan School, education economics

The methodological innovation at the heart of the study lies in what the authors call health impact event attribution. Traditional extreme event attribution, which has matured rapidly over the past decade, typically asks how climate change altered the probability of a meteorological anomaly such as a heatwave, using metrics like regional temperature maxima. But a hotter day does not translate directly into a death, and the relationship between temperature and mortality varies enormously with geography, climate, age structure and adaptation. The new approach closes that gap by coupling attribution statistics to epidemiological exposure-response functions calibrated for each of 34 European countries. Rather than asking how much more likely the 2022 heatwave was, the team asked how much more likely the mortality event itself was, the actual human toll, in a climate warmed by roughly 1.2 degrees Celsius compared with a pre-industrial baseline.

Technically, the researchers constructed generalized extreme value distributions for both temperature and heat-related mortality, allowing them to estimate return periods, how often an event of a given magnitude would be expected to occur, under current and counterfactual pre-industrial conditions. The ratio of these probabilities, known as the probability ratio, then expresses how much climate change has multiplied the odds of an extreme. Crucially, the mortality side of the analysis was built on weekly death counts from Eurostat, publicly available data that are updated in near real time. Temperature inputs came from the ERA5-Land reanalysis produced by the European Centre for Medium-Range Weather Forecasts, while global mean surface temperature changes were drawn from NASA’s GISTEMP record. The reliance on weekly rather than daily mortality data was validated using recent methodological work showing that temporally aggregated health data can yield unbiased temperature-mortality estimates, a finding that dramatically expands where such analyses can be performed.

The results reveal a continent divided. Across Europe as a whole, 2022-like heat-mortality events became substantially more probable, but the amplification in Southern Europe dwarfed the continental average. The Mediterranean region, already identified in the scientific literature as a climate change hotspot, experienced nonlinear increases in the likelihood of deadly heat, meaning that each additional fraction of a degree of warming multiplies the risk faster than the last. The study also projected forward, estimating probability ratios at 1.5 and 2.0 degrees Celsius of global warming, the benchmark thresholds of the Paris Agreement. Those projections show the risk continuing to climb steeply, with Southern Europe again bearing the steepest increases. The nonlinear character of these changes is particularly troubling because it implies that modest additional warming will produce disproportionately large jumps in mortality risk, compressing the time available for societies to adapt.

Equally significant are the demographic disparities the analysis uncovered. When the researchers stratified their models by age, they found that the elderly, particularly those aged 80 and older, faced sharply elevated probability ratios compared with younger populations. This pattern reflects well-documented physiological vulnerabilities, including diminished thermoregulatory capacity, higher prevalence of chronic disease, and greater use of medications that impair the body’s ability to cope with heat stress. Prior research has also documented sex-specific differences in heat vulnerability and adaptation, and the new framework’s ability to dissect such heterogeneity is one of its principal strengths. The authors emphasize that these findings make the case for population-specific analyses: a national average can conceal the fact that a small, highly vulnerable subgroup absorbs a wildly disproportionate share of the climate-driven mortality increase.

The 2022 summer serves as the study’s anchor case because it was, by most measures, the hottest summer ever recorded in Europe. Spain registered its warmest summer in its historical series, the United Kingdom breached 40 degrees Celsius for the first time, and the Copernicus Climate Change Service documented the season as a continental anomaly. Previous work by several of the same authors estimated that the summer of 2022 caused more than 60,000 heat-related deaths across Europe, while companion studies attributed a substantial fraction of that toll directly to anthropogenic warming. The new attribution analysis reframes those mortality counts in probabilistic terms: without human influence on the climate, an event of that lethality would have been extraordinarily rare, whereas in today’s climate it has become a recurring hazard with a return period measured in years rather than centuries.

What distinguishes this work from earlier impact attribution efforts is not only its scope, spanning 34 countries, but its speed and reproducibility. Because the mortality data are public and updated weekly, and because the analytical code has been released openly on GitHub, the framework can in principle be deployed rapidly after any extreme heat episode, transforming attribution from a retrospective academic exercise into a near-real-time public health tool. The authors argue that this capability could feed directly into health emergency forecasting systems, several of which are already being developed and tested in Europe, allowing authorities to anticipate and respond to deadly heat before the peak toll is counted. Rapid attribution of health impacts also carries weight beyond epidemiology, providing quantified evidence relevant to climate litigation, loss-and-damage negotiations and the legal and moral accounting of emissions.

The study arrives amid a broader scientific consensus that heat has become Europe’s leading weather-related killer. The Lancet Countdown and the World Meteorological Organization have both documented rising heat mortality, and modeling studies covering hundreds of European cities project that the burden will grow substantially without aggressive adaptation. Yet adaptation itself remains uneven. Research from Spain and the Netherlands shows that populations can shift their minimum mortality temperature over time, a signature of acclimatization, but the pace of such physiological and behavioral adaptation lags far behind the pace of warming. Heat-health warning systems exist across much of Europe but vary widely in coverage, trigger thresholds and effectiveness. The new findings sharpen the argument that adaptation investments, from cooling centers and urban greening to occupational heat protections and targeted outreach to the elderly, should be prioritized precisely where and for whom the probability ratios are highest.

Ultimately, the study delivers a double message. First, the fingerprint of climate change is now legible not just in thermometers but in mortality statistics: the deaths of thousands of Europeans in recent summers are statistically attributable to emissions, with Southern Europeans facing risks an order of magnitude beyond the continental average. Second, the tools needed to see that fingerprint are now fast, transparent and grounded in freely available data, meaning that the era of waiting years to understand the human cost of an extreme summer is over. As global temperatures continue to rise toward and potentially past the 1.5-degree threshold, the nonlinear escalation of heat-mortality risk documented in this analysis offers both a warning and a roadmap. The regions and populations identified as most vulnerable are known, the data streams needed to monitor the threat in real time exist, and the remaining variable is the speed with which societies choose to act on the evidence.

Subject of Research: Attribution of extreme heat-related mortality events in Europe to anthropogenic climate change

Article Title: Extreme event attribution for heat-related mortality due to anthropogenic climate change across Europe

Article References: Beck, T. M., Gudmundsson, L., Schumacher, D. L., Seneviratne, S. I., Achebak, H., & Ballester, J. (2026). Extreme event attribution for heat-related mortality due to anthropogenic climate change across Europe. Nature Health. https://doi.org/10.1038/s44360-026-00193-z

Image Credits: AI Generated

DOI: 10.1038/s44360-026-00193-z

Keywords: extreme event attribution, heat-related mortality, climate change, Europe, heatwaves, Nature Health, epidemiological models, Southern Europe, public health, adaptation, probability ratio, 2022 heatwave