Although the researchers were able to demonstrate the ability of LLMs to extract relevant details about GI symptoms and foods eaten, and the researchers believe in the potential of an AI surveillance approach for aiding future foodborne illness investigations, several significant challenges were identified and must be overcome before the technology can be effective.

Specifically, the imperfect nature of the data itself presents a potential source of bias, as only a subset of all food consumed is prepared outside the home and would therefore generate an online review. However, the researchers found no evidence that the LLM’s text processing would introduce any new bias.

Another challenge is the onset of foodborne illness symptoms ranging from hours to days, which may lead reviewers to misattribute their sickness to a foodservice establishment that is not actually the source. Variations in spelling and the use of slang were also identified as obstacles.

The researchers identified the biggest limitation to be the difficulty in determining which specific ingredients are used in the foods that ill reviewers attribute to their symptoms.

Although further work is needed before UKHSA can adopt AI surveillance of online reviews as a routine foodborne illness outbreak investigation activity, the agency is optimistic that AI could soon help identify potential sources of future foodborne illness outbreaks, when used in combination with traditional epidemiological tools.