Comparing researcher-curated AI-drafted feedback styles for negative automatic thought believability: an online...

AI-drafted feedback is increasingly used in learning support, emotional support, peer support, and self-guided reflection, yet empirical studies often combine emotional validation, advice, reappraisal, and Socratic questioning within the same response. This study compared four researcher-curated AI-drafted feedback styles: neutral information, empathetic feedback, Socratic challenge, and integrated feedback.

An online randomized scenario-based experiment was conducted using four daily stress scenarios crossed with the four feedback conditions. After screening 427 questionnaire records collected from March 12 to March 20, 2026, valid responses from 389 adults were analyzed. Participants read one scenario, rated negative automatic thought (NAT) believability and negative affect, received one feedback text, completed post-test and feedback evaluation measures, and wrote a balanced alternative thought. Regression models controlled for pre-test scores, scenario, trait NAT tendency, AI trust, and trait reactance.

All three active feedback styles predicted lower post-test NAT believability than neutral information, with the largest adjusted difference observed for integrated feedback, followed by Socratic challenge and empathetic feedback. Empathetic feedback produced higher perceived empathy and greater affect reduction, whereas Socratic challenge produced higher cognitive reappraisal and reframing quality. Integrated feedback combined high feedback acceptance with strong evidence-based examination of the negative thought.

In this scenario-based task, researcher-curated AI-drafted feedback produced larger reductions in NAT believability and higher evidence-based reframing quality when emotional validation was followed by evidence testing, alternative explanation, and fact-inference distinction.