Large Language Models and Pseudo-Information

Interesting conclusions from Schmitz (2025) in a recent Health Information and Libraries Journal (HILJ) article* on comparative research into using LLMs to identify erroneous health information. With ongoing concerns about the impact of AI on misinformation and disinformation, this informative piece provides further evidence of the need for ongoing, rigorous research in the context of valid, reliable and trustworthy journal content.

The premise was to identify to what extent LLMs (specifically ChatGPT, Claude 3.5 Sonnet, Gemini and Copilot) can be used to flag pseudo-information when presented with an open-access article, influenced by the actions of Tom Spears, a reporter who famously smuggled in pseudo-scientific articles to open-access research to highlight the "ridiculous behaviour of pseudo-scientific journals" (Schmitz, p.2).

Key takeaways from this;

Open access science is an important source of health information, but erroneous or pseudo-information is harmful
LLMs have the potential to improve access to health-related content
LLMs can identify pseudo-information accurately, to some degree, though the reproducibility of the research is low
Currently, LLMs cannot replace critical thinking in identifying pseudo-information
Users in health knowledge services should be offered training on LLMs

With such an in-depth review, the author concludes:

"It seems necessary that government agencies such as the National Health Service (NHS) Library and Knowledge Services issues guidelines on how AI-based support systems, including LLMs can be used to assist health information and library services workers as well as health and public health workforce" (Schmitz, p.9).

Furthermore, there is a need for more testing and validation of AI tools and mandatory training programmes for users in health knowledge services, supported by regular workshops and conferences.