Causal reasoning about epidemiological associations in conversational AI

Area of Science:

• Epidemiology
• Artificial Intelligence
• Public Health

Background:

• Epidemiological studies frequently report associations between fine particulate matter (PM2.5) and increased human mortality risks.
• Large language models (LLMs) like ChatGPT can reflect human reasoning patterns found in their training data.
• The causal interpretation of environmental health data is critical for public health policy.

Purpose of the Study:

• To explore the causal interpretation of epidemiological associations between PM2.5 and mortality risks using a Socratic dialogue with ChatGPT.
• To assess the LLM's ability to critically evaluate scientific evidence and refine its conclusions through questioning.
• To evaluate the potential of LLMs in improving scientific reasoning and the reliability of initial conclusions.

Main Methods:

• A Socratic dialogue methodology was employed, engaging ChatGPT in a question-and-answer format.
• The dialogue focused on the interpretation of evidence linking PM2.5 exposure to human mortality.
• ChatGPT's responses were analyzed for shifts in reasoning and conclusion certainty.

Main Results:

• Initially, ChatGPT asserted the established link between PM2.5 and mortality, emphasizing public health importance.
• Through sustained questioning, ChatGPT revised its stance, concluding that causality is uncertain due to potential omitted confounders.
• This demonstrates an LLM's capacity to refine its interpretation of evidence when prompted.

Conclusions:

• Sustained Socratic questioning can improve the reasoning and argumentation imitated by LLMs.
• LLMs may initially present strong conclusions that require further interrogation for nuanced understanding.
• The reliability of LLM-generated conclusions in scientific interpretation can be enhanced through interactive dialogue.