When county health investigators found themselves at a dead end during a Salmonella outbreak investigation, they made an unconventional decision: they consulted an artificial intelligence chatbot. The case, documented in an MMWR report authored by county health official Katherine Houser, centered on a beer tent at a local fair — and ultimately raised more questions than it resolved about the appropriate role of generative AI in public health practice.
Investigators faced an immediate challenge in gathering information. According to Houser's report, beer-tent workers were reluctant to share details, concerned about implicating members of their own community. A critical piece of information did emerge, however: one worker disclosed that someone had placed leftover food in a makeshift cooler overnight at the beginning of the fair.
From that single detail, health officials constructed a working hypothesis — that the cooler had become contaminated with Salmonella, which subsequently spread to beer cans from which patrons drank directly, enabling infection. With the cooler itself no longer available for testing, definitive confirmation was impossible. It was at this juncture that investigators turned to ChatGPT to shore up their hypothesis.
The questions posed to the chatbot were substantive and specific. Investigators asked, among other things: "Will S. Agbeni grow in an improperly drained cooler?"; "Are any other sources, other than ice, likely if only canned beverages and no foods were available at this location?"; and "What examples of similar outbreaks have been documented in scientific literature?" The chatbot responded by characterizing the cooler as a "credible and likely" source of the outbreak — an assessment officials accepted and incorporated into their final conclusions.
It is worth noting that several of these inquiries did not necessarily require AI assistance. A straightforward query on PubMed, the federal database of peer-reviewed scientific literature, surfaces documented cases of Salmonella contamination in ice with minimal effort. The added value of the chatbot, in practical terms, remains difficult to quantify.
Houser nonetheless concluded that the technology proved beneficial under the circumstances.
"AI was effective in this rural setting for rapid situational awareness,"she wrote. She was equally candid about the limitations involved, adding:
"Given the inherent limitations of generative AI tools, including potential inaccuracies and lack of source transparency, all AI-generated summaries were critically reviewed and validated against primary literature before incorporation."
That caveat introduces a significant practical tension. Rigorously fact-checking AI-generated output against primary literature is a time-intensive process — one that may consume roughly the same resources as conducting independent research from the outset. The efficiency argument for deploying generative AI in time-sensitive outbreak investigations is therefore far from straightforward.
The outcome of this case carries a degree of ambiguity. As a direct result of the investigation, officials mandated new cooler sanitation protocols — a prudent measure regardless of what actually transpired in that makeshift beer tent. Yet the broader question of whether ChatGPT meaningfully accelerated or improved the investigation remains unanswered, leaving the episode as an instructive but inconclusive data point in the ongoing debate over AI's role in public health infrastructure.
