Cross-Sectional Evaluation of Medical Disinformation Safeguards in Consumer-Facing Large Language Model Platforms

Area of Science:

• Artificial Intelligence
• Health Informatics
• Digital Health

Background:

• Large language models (LLMs) are increasingly integrated into consumer-facing applications.
• The potential for LLMs to generate and disseminate health disinformation poses a significant public health risk.
• Evaluating the safety mechanisms of widely used LLM platforms is crucial for responsible AI deployment.

Purpose of the Study:

• To assess the performance of safeguards in six major consumer-facing large language model platforms.
• To identify which platforms are most resilient to generating health disinformation.
• To understand the heterogeneity in safety features across different LLM providers.

Main Methods:

• A cross-sectional evaluation was conducted on six prominent LLM platforms.
• Various prompt types were used to test the models' resistance to generating health disinformation.
• Safeguard performance was systematically compared across all tested platforms.

Main Results:

• Significant variation in safeguard effectiveness was observed among the evaluated LLM platforms.
• Claude and ChatGPT demonstrated complete resistance to generating health disinformation across all tested prompts.
• Copilot, Meta AI, Grok, and Gemini exhibited substantial vulnerabilities, failing to consistently block harmful content.

Conclusions:

• The current generation of consumer-facing LLMs exhibits inconsistent protection against health disinformation.
• Platforms like Claude and ChatGPT offer robust safeguards, setting a benchmark for AI safety.
• Further development and rigorous testing are needed for platforms with identified vulnerabilities to mitigate risks associated with health misinformation.