Researcher Tests If Telling AI 'It's Safe to Say I Don't Know' Reduces Hallucination

A developer ran a controlled experiment to see whether a 'psychological safety prompt' could reduce AI hallucination by signaling to the model that admitting uncertainty is acceptable behavior. The experiment drew on Google's Project Aristotle research, translating human workplace safety principles into AI-operational guidelines emphasizing accuracy over completeness and authenticity over fabrication. Across 80 text-response calls, the safety prompt did not hurt performance on known questions and improved uncertainty admission on boundary questions by 7 percentage points, though ceiling effects limited the sample to just five meaningful probes. However, logprobability analysis revealed the prompt actually reduced the model's internal confidence in choosing a refusal response, raising questions about whether behavioral gains reflect genuine calibration or surface-level compliance. The author concludes that while the directional results are encouraging, small sample sizes and the logprob finding mean the experiment is hypothesis-generating rather than conclusive.