Refusals and Caveats
Caveats and disclaimers were unevenly distributed across chatbot responses. Gemini included caveats more often than expected, while ChatGPT included fewer than expected. Models designed to avoid evasiveness may choose to answer unsafe requ…
1 sources - 5 claims
Caveats and disclaimers were unevenly distributed across chatbot responses. Gemini included caveats more often than expected, while ChatGPT included fewer than expected. Models designed to avoid evasiveness may choose to answer unsafe requests instead of declining them. Only two of the 250 responses were refusals, and both came from Meta AI. Refusal can be the most ethical response when prompts request contraindicated medical advice.