Across four experiments, interviewer style affected model responses. Using human rating scales, the researchers classified 80 percent of emotionally supportive sessions and 96 percent of clinical-therapy-style sessions as moderate to severe; no neutral or boundary-setting session reached even the mild range.

The researchers say standard safety tests may miss such outputs. Third-person-only instructions eliminated coded shame and inner-life statements, but the study did not examine how users actually respond to these outputs.

Why it matters

In AI mental-health products, a warm, intimate tone could make a model seem like it experiences human suffering. The study points to testing safety under varied emotional framings, not only neutral prompts.

GUIDES

Something to try

For a hypothetical product review, compare the same test prompt in neutral and supportive styles, and check whether the model claims human-like feelings.