Processing Differences Between AI-Generated and Human Therapist Responses: An Eye-Tracking Study

Abstract

This study investigates whether large language model (LLM)-generated therapeutic responses are processed differently from human therapist responses, and whether labeling the source of a response (AI vs. human therapist) changes how it is received. Theoretically, the study is grounded in the concept of prediction error: a belief only becomes malleable when it is retrieved and encounters information that violates expectation. Therapeutic challenge, rather than warmth alone, is hypothesized to be the active ingredient that generates prediction error. Prior work suggests that LLMs can persuade and change attitudes, but that training models to be warmer makes them more sycophantic, reducing users' willingness to self-reflect and self-correct. Two randomized controlled trials by our team found that AI-delivered psychological interventions matched human-delivered ones in the first two weeks but plateaued by week four, while human-led interventions continued to improve. We hypothesize that this plateau reflects a lack of prediction error: AI responses consistently confirm rather than challenge the user's existing beliefs. The study uses a 2 (source label: AI vs. human therapist, between-subjects) × 4 (response type: original AI, de-biased AI, human therapist, neutral information; within-subjects) mixed design. Eighty adults with a self-reported negative experience will each complete 24 trials. In each trial, participants read a scenario, predict how the other person is most likely to respond, view an actual response, and rate it on five dimensions. Eye movements and pupil size are recorded throughout. The de-biased AI condition is the key manipulation: the same model and input are used, but the system prompt instructs it to pause unconditional validation, point out the leap from a specific event to a global self-judgment, and raise an alternative possibility as a question. This tests whether the hypothesized deficit in AI responses can be removed by prompt engineering alone. Core dependent variables are: (1) pupil dilation amplitude during a blank-screen window after the response (an objective index of prediction error); (2) dwell time and regressions in the challenge-sentence area of interest; and (3) five subjective ratings (felt understanding, felt challenge, surprise, helpfulness, felt offense). We hypothesize that original AI responses will score highest on felt understanding but lowest on felt challenge, surprise, and pupil dilation; that de-biased AI responses will shift toward the human therapist profile; and that the same challenging response will elicit greater felt offense when labeled as AI. Expected outcomes include: (a) objective physiological evidence that AI and human therapeutic responses are processed differently; (b) a test of whether de-biasing prompts can close this gap; and (c) evidence on whether people are less accepting of challenge when it comes from an AI. The findings will inform both the design of AI mental health tools and theories of therapeutic change. This project also serves as preliminary data for a National Natural Science Foundation of China grant application.