HHS Developing AI Tool to Formulate Hypotheses on Vaccine Injury Claims

The US Department of Health and Human Services is creating a generative artificial intelligence tool designed to identify patterns in data reported to a national vaccine monitoring database and generate hypotheses regarding the potential negative effects of vaccines, as outlined in a recently released inventory detailing the agency’s AI use cases for 2025.

As stated in the HHS document, the tool has not been implemented yet, and an AI inventory report from the previous year indicates that it has been in development since late 2023. However, experts express concerns that the predictions produced could be exploited by Health and Human Services Secretary Robert F. Kennedy Jr. to advance his anti-vaccine agenda.

Kennedy, a long-time vaccine critic, has altered the childhood vaccination schedule during his tenure, removing several recommended immunizations for children, including those for Covid-19, influenza, hepatitis A and B, meningococcal disease, rotavirus, and respiratory syncytial virus (RSV).

Additionally, Kennedy has advocated for reforming the current safety monitoring system for vaccine injury data collection, known as the Vaccine Adverse Event Reporting System (VAERS), arguing that it conceals information regarding the actual rate of vaccine side effects. He has also proposed modifications to the federal Vaccine Injury Compensation Program that could facilitate lawsuits for adverse events not established as linked to vaccines.

VAERS, jointly overseen by the Centers for Disease Control and Prevention and the Food and Drug Administration, was created in 1990 to identify potential safety concerns with vaccines post-approval. Reports of adverse reactions can be submitted by anyone, including healthcare providers and the public. However, since these claims are unverified, VAERS data alone is insufficient to ascertain whether a vaccine resulted in an adverse event.

“At best, VAERS has always been a mechanism for generating hypotheses,” states Paul Offit, a pediatrician and director of the Vaccine Education Center at Children’s Hospital of Philadelphia, who previously served on the CDC’s Advisory Council on Immunization Practices. “It’s a noisy system. Anyone can report, and there’s no control group.”

Offit notes that the system only reflects adverse events that occur sometime after immunization and does not demonstrate that a vaccine caused those reactions. The CDC’s own website clarifies that a report to VAERS does not imply that a vaccine caused an adverse event. Despite this, anti-vaccine advocates have historically misused VAERS data to claim vaccines are unsafe.

Leslie Lenert, former founding director of the CDC’s National Center for Public Health Informatics, mentions that government scientists have utilized traditional natural language processing AI models to analyze VAERS data for years, making HHS’s move toward advanced large language models unsurprising.

One significant drawback of VAERS is its lack of data regarding how many people received a vaccine, which can lead to an overrepresentation of logged events. For this reason, Lenert emphasizes the necessity of combining VAERS information with other data sources to accurately assess the true risk of an event.

Moreover, large language models are known for generating plausible but inaccurate outputs, highlighting the importance of human follow-up on any hypotheses produced by an LLM.

“VAERS is meant to be exploratory. Some individuals in the FDA are now perceiving it as more than just exploratory,” notes Lenert, who is currently the director of the Center for Biomedical Informatics and Health Artificial Intelligence at Rutgers University.