The US Department of Health and Human Services is building a generative artificial intelligence tool intended to examine reports in a national vaccine monitoring database and produce hypotheses about negative vaccine effects, according to an inventory of the agency’s AI use cases for 2025.
The tool has not been deployed, the HHS document says. A prior AI inventory report shows it has been in development since late 2023. The project is attracting attention because of what the system may generate, how those outputs could be interpreted, and the political context around vaccine safety oversight.
What HHS says the tool is meant to do
The planned system is designed to find patterns across data reported to the Vaccine Adverse Event Reporting System, or VAERS. It would use generative AI to create hypotheses about possible adverse effects associated with vaccines.
That wording matters. A hypothesis is not a conclusion. It is a lead that may point researchers toward a question worth studying. In vaccine safety, that distinction is central because VAERS is an open reporting system, not a database of verified vaccine-caused injuries.
VAERS is jointly managed by the Centers for Disease Control and Prevention and the Food and Drug Administration. It was established in 1990 to help detect potential vaccine safety issues after approval. Reports can be submitted by health care providers and by members of the public.
Because reports are not verified before they enter the database, VAERS data alone cannot show that a vaccine caused a health event. The system records adverse events that happened sometime after vaccination, but that timing does not prove causation.
Why VAERS is powerful but limited
Experts quoted in the source article describe VAERS as a useful early-warning tool, but one that must be handled carefully. Paul Offit, a pediatrician and director of the Vaccine Education Center at Children's Hospital of Philadelphia who was previously a member of the CDC’s Advisory Council on Immunization Practices, described VAERS this way:
“VAERS, at best, was always a hypothesis-generating mechanism,” says Paul Offit. “It's a noisy system. Anybody can report, and there’s no control group.”
That means VAERS can help surface signals, but it cannot answer the full safety question on its own. One major limitation noted by Leslie Lenert, previously the founding director of the CDC’s National Center for Public Health Informatics, is that VAERS does not include data on how many people received a vaccine. Without that context, events in the database can appear more common than they actually are.
Lenert says VAERS information should be paired with other data sources to determine the true risk of an event. That need becomes even more important if a large language model is used to sift through the reports, because LLMs are known for generating convincing hallucinations.
In plain terms, an AI system may produce a pattern or explanation that sounds plausible but still needs scientific validation. The useful output is the question it raises, not the answer it appears to provide.
The political stakes around vaccine safety data
The tool is being developed while Health and Human Services secretary Robert F. Kennedy Jr., a long-standing vaccine critic, has made major changes to vaccine policy in his year in office. The source article reports that Kennedy has removed several shots from the list of recommended immunizations for all children, including those for Covid-19, influenza, hepatitis A and B, meningococcal disease, rotavirus, and respiratory syncytial virus, or RSV.
Kennedy has also called for overhauling the safety monitoring system used for vaccine injury data collection. He has claimed that VAERS suppresses information about the true rate of vaccine side effects. He has also proposed changes to the federal Vaccine Injury Compensation Program that could make it easier for people to sue for adverse events that have not been proven to be associated with vaccines.
That background is why experts are watching the AI project closely. An AI tool that produces hypotheses from VAERS could be useful if its findings are treated as leads for further study. It could also be risky if preliminary outputs are treated as proof.
Lenert, now the director of the Center for Biomedical Informatics and Health Artificial Intelligence at Rutgers University, says government scientists have used traditional natural language processing AI models to search VAERS data for patterns for several years. In that sense, moving toward more advanced large language models is not surprising. The concern is how the results are interpreted and used.
False alerts and the need for expert review
Jesse Goodman, an infectious disease physician and professor of medicine at Georgetown University, says large language models could potentially help detect previously unknown vaccine safety issues. But he also warns that VAERS can contain inaccurate and incomplete data, so any leads need thorough investigation.
“I would expect, depending on the approaches used, a lot of false alerts and a need for a lot of skilled human follow-through by people who understand vaccines and possible adverse events, as well as statistics, epidemiology, and challenges with LLM output,” he says.
Goodman also points to deep staffing cuts at the CDC as a practical concern. If an AI system produces new signals, agencies need the plans and capacity to screen the data, decide what deserves further study, and determine how that study should be done.
The source article notes that VAERS has identified real safety concerns before. It flagged instances of a rare clotting disorder among some people who received the Johnson & Johnson Covid-19 vaccine and rare cases of myocarditis, particularly among younger males, who received mRNA Covid-19 vaccines.
Those examples show why the database exists. They also show why follow-up matters. VAERS can point investigators toward a possible problem, but the signal must still be tested with expertise, additional data, and appropriate methods.
What happens next
For now, the HHS generative AI vaccine injury tool remains undeployed, according to the agency document described in the source article. Its promise is straightforward: faster pattern detection across a large reporting system. Its risk is equally clear: preliminary hypotheses could be mistaken for established evidence.
The debate is not simply about whether AI belongs in vaccine safety monitoring. Traditional AI methods are already part of pattern-finding work, according to Lenert. The harder question is whether a generative AI system can be governed in a way that keeps its outputs exploratory, transparent, and subject to human scientific review.
HHS did not respond to a request for comment.