Public Health Institutions Face AI-Flooded Information and Defunded Evidence Systems

Public health institutions are losing staff, funding and independent authority at the same time that generative artificial intelligence is flooding their working environment with fluent, often unsourced information. Kent Buse and Sarah Hawkes make this case in an opinion article published in PLOS Global Public Health, arguing that distinguishing measured evidence from unsupported claims has become a core responsibility of public health institutions. Their argument is not that verification systems have disappeared, but that the organisations responsible for producing and protecting reliable evidence are being weakened while AI systems make plausible statements, fabricated citations and model-generated estimates easier to produce. The article concerns public health evidence systems broadly, including international organisations, national agencies, statistical offices, universities and research institutes, rather than reporting a controlled study or evaluating the performance of a specific AI product.

The clearest example in the article is the United States decision in February 2025 to end funding for the Demographic and Health Surveys programme. For roughly four decades, the programme produced population trends and estimates of child mortality, nutrition, immunisation and other health conditions used by health ministries, United Nations agencies and researchers in more than 90 countries. Surveys were under way or ready in 23 countries when the funding stopped. The annual cost had been about US$47 million, less than 0.1% of the USAID budget, and private philanthropy subsequently funded a reduced programme. In the authors’ assessment, a time-limited grant now supports a backbone of population-health data used across much of the world, exposing the fragility of evidence infrastructure when public institutions depend on a single donor or foundation.

Generative AI is already entering the institutions expected to perform this verification. By late 2024, the United Nations system had catalogued more than 700 AI projects across some 50 entities, almost half of them involving text and language. The World Health Organization launched a generative health chatbot in 2024, and an early assessment found operational failures that risked the organisation’s credibility rather than strengthening it. The authors also describe checking the status of the Demographic and Health Surveys programme while writing the article; the first answer they received was an AI-generated summary that was confident and only partly sourced. Their example does not establish that every generated answer is inaccurate, but it illustrates how quickly a statement can appear without revealing whether it was measured, modelled, quoted or fabricated.

Buse and Hawkes describe information management as a question of power rather than merely an information-technology problem. How WHO, UNICEF, national public health institutes and statistics offices manage information helps determine whose evidence is admitted, which issues reach policy agendas and what publics and decision-makers eventually regard as common sense. Drawing on theories of power associated with Bachrach, Baratz and Lukes, they distinguish visible decisions, control of the agenda and the less visible shaping of preferences. The third is especially important to their argument because its effects can remain unnoticed: one of the clearest signs that preferences are being shaped is that a question is never asked. Redirecting what counts as evidence and discrediting those who certify it are therefore institutional actions even when they do not take the form of an explicit order to suppress a fact.

The article notes that such strategies are not new. Tobacco companies used their power to manufacture doubt for more than half a century, with doubt needing only to persist rather than prevail in an argument. The sugar industry funded research in the 1960s that redirected scientific attention from sugar to fat, shaping dietary guidance for a generation. Corporate lobbying of United States positions toward WHO averaged US$12.5 million per year between 2006 and 2015, and the authors say that part of its apparent purpose was to erode confidence in the institution. What has changed, they argue, is the cost, speed and reach of the machinery now available. Generative AI can produce large volumes of plausible text with unprecedented efficiency, allowing institutions and actors to shape the information environment at a scale that was previously difficult to sustain.

To explain the technical challenge, the authors separate information, evidence, knowledge and fact. Information includes whatever moves through the environment, such as data, text, testimony, rumour and fabrication. Evidence is the subset of information that bears on the truth of a claim and can justify action. Knowledge is belief that is both true and justified, while a fact is what makes a true claim true whether or not anyone believes it. Public institutions exist in part to move material through this chain, sifting information into evidence and evidence into knowledge that can guide decisions. Sifting is slow, contestable and expensive because it rests on epistemic authority: the socially recognised standing of institutions and experts to determine what counts. That authority cannot be recreated merely by producing a confident answer on demand.

Large language models act throughout this chain by generating abundance without discernment. They can place a peer-reviewed estimate, a modelled projection, a figure taken from a press release and a hallucination in the same fluent register, removing the provenance needed to distinguish them. They can also fabricate citations at rates that, in conventional research, would end a career. Their training corpora also reproduce the prevailing distribution of published knowledge, including its emphasis on English, the Global North and conflicts of interest that can shape findings. The article connects this development to Herbert Simon’s 1971 observation that a wealth of information creates a poverty of attention: as more information becomes available, less of it can be assessed. WHO named the pandemic version of this problem an infodemic, and the authors argue that AI models industrialise the problem rather than resolve it.

Some of the pressure on evidence systems predates generative AI. Around half the health-related Sustainable Development Goal indicators are now tracked primarily through modelled estimates rather than data reported by countries, partly because many countries lack the capacity and resources for ongoing health surveillance. Modelled estimates are information, and identifying the evidence within them has traditionally been a responsibility of scientists rather than policy-makers. The authors argue that scientists, statisticians and an independent press face sustained attacks at a time when actors benefiting from an evidence-free process are also discrediting the institutions that certify evidence. In that setting, policy-makers are increasingly left to resolve the distinction between information and evidence themselves. The source frames this as a political and institutional development, not as proof that modelled estimates are themselves invalid.

The authors argue that this work cannot be left to AI platforms, commercial vendors or individual users. Model developers sell fluency rather than verification, and the discernment required at the scale of a health system remains an institutional function. Public health agencies, research institutes including universities, statistical offices and multilateral organisations have both a public mandate and, for the time being, the legitimacy to exercise this authority. These institutions already possess relevant structures: WHO has a chief scientist, a guidelines review committee and a science council; UNICEF has a chief statistician and research office; and the UN system has a chief statistician. However, the authors say these functions were designed to certify what an institution publishes, not to govern how thousands of staff use commercial AI tools to generate text and estimates, even as the existing functions are being reduced.

Institutional retrenchment makes the problem more difficult. The article reports that WHO’s 2025 restructuring reduced director posts from 76 to around 34, with up to 30% to 40% of headquarters staff facing redundancy. Buse and Hawkes propose a response with a technical and a political dimension. Technically, agencies need provenance standards that show whether a figure was measured, modelled or generated, together with its uncertainty; evidence functions capable of working at the speed of politics without abandoning verification; renewed investment in primary data collection and national statistical capacity; and diversified financing so that a survey programme costing less than 0.1% of one donor’s aid budget does not depend on that donor or a single foundation. Politically, institutions must defend independent, transparent and accountable scientists and statisticians. The authors add that silence when those workers are attacked is itself an information-management decision. Their conclusion is conditional: institutions that learn to curate, contest and defend knowledge as a public good may remain trusted over the coming decade, whereas those that outsource the question of what is true to whoever owns a model risk losing both a distinctive public function and the basis of their legitimacy.

Subject of Research: Public health evidence institutions, generative AI and the verification of policy-relevant information

Article Title: Who sifts? Knowledge as a public good in the age of AI

Article References: Buse, K., & Hawkes, S. (2026). Who sifts? Knowledge as a public good in the age of AI. PLOS Global Public Health, 6(10), e0007381. https://doi.org/10.1371/journal.pgph.0007381

Image Credits: AI Generated

DOI: 10.1371/journal.pgph.0007381

Keywords: generative AI, public health evidence, knowledge verification, epistemic authority, health statistics, Demographic and Health Surveys, WHO, public institutions, information overload, provenance, sifts, Knowledge