Trial of 380 adults shows AI reassurance and coaching soothe equally, but only coaching helped later

The differences came afterwards. The coaching group reported higher self-efficacy (b = 0.500), had better blind-coded written strategies (b = 0.353) and recovered more during a later task without AI (0.99 vs 0.71 points). The reassurance group more strongly believed they needed the AI for comparable relief (b = 1.010) and showed stronger intentions to rely on it (b = 0.628).

The results suggest that judging AI companions only by how much better users feel in the moment misses differences in what users take away. The authors stress the reliance findings are short-term beliefs and intentions, not emotional dependence. In an exploratory choice, 36.8% of the reassurance group asked for another AI reply versus 27.9% of the coaching group, a difference that was not statistically conclusive (p = 0.063).

The experiment was a two-condition randomized online study with concealed permuted-block allocation. After a mild autobiographical emotion induction, each participant had three exchanges with a GPT-5.5-based companion given one of two system prompts. Pilot testing with 80 other adults confirmed the two styles were comparable in warmth, empathy and helpfulness, and transcript coding found 95.3% and 95.8% of replies met the full fidelity threshold.

About three minutes later, participants faced a second standardized setback with no AI and wrote how they would regulate their response. Affective recovery during that task was 0.99 points after coaching versus 0.71 after reassurance (adjusted b = 0.279). All five focal effects survived Holm correction, and fidelity-restricted and expanded-covariate analyses changed coefficients by at most 0.015.

The reassurance prompt told the AI to emphasize continued availability and encourage returning for comfort. It did not claim consciousness or exclusivity and did not discourage human support. Participants in that group rated the AI as more necessary and said they would rely on it more, yet in the exploratory request-another-AI choice the gap was inconclusive (OR = 1.51, 95% CI 0.98 to 2.32).

The authors say the findings point to a design goal of warm but bounded support: validation combined with emotion labeling, limited reappraisal and a user-chosen action, while scrutinizing response policies that repeatedly position the AI as the source of calm.