Differences in health advice from AI vary by access method
Challenges of Auditing: Variability in Outputs of Large Language Models for Health
Computation and Language
Summary
People ask AI models for health advice through different ways like chatbots or APIs, but these different methods give different answers. The authors found that using APIs to test these models might not match what people get when they use chatbot interfaces. This means you can't fully trust current tests to show how well the AI actually works for most users. The authors say companies need to let auditors copy the exact user experience to check the AI properly.
What this means in practice
- •For healthcare technology teams: Test AI health tools across all user access modes to ensure consistent advice for patients and providers.
- •For ai compliance officers: Design audits that match real user experiences by replicating the exact interface and settings AI consumers use.
Authors
Yuan Pu, Yewon Chang, Furong Jia, Xunjian Yin, Jessica Ma, Ayman Ali, Monica Agrawal
Abstract
People increasingly use frontier AI models for health advice, but via different access modes (e.g., ChatGPT, ChatGPT Health, APIs) with varying settings. Here, we find systematic differences across access modes. Because evaluations typically rely on APIs while consumers interact through chatbot interfaces, these discrepancies limit evaluation validity. Our findings underscore an urgent need for model providers to enable faithful replication of consumer experiences and settings for rigorous audits.