Generative AI-assisted self-visualization and self-concept clarity: evidence consistent with a cognitive-mirroring mechanism in human–AI interaction

Generative artificial intelligence (AI) now lets people turn verbal self-descriptions into visual self-representations, creating a new setting for studying self-concept clarity (SCC) in human–AI interaction. We tested whether generative AI-assisted self-visualization (GAISV) was associated with higher short-term SCC through cognitive mirroring rather than through general immersive flow, and whether this depends on how far participants identify with the generated image. We introduce the AI Mirror Self-Integration Model and evaluated it in two stages. A pilot study (124 valid participants) examined the Cognitive Mirroring in Human-AI Self-Visualization Scale (CM-HAIS) and supported its use as a task-state measure of reflective self-processing. A randomized main experiment then recruited 328 valid participants. Primary analyses compared three conditions—GAISV, active non-self image generation, and passive image viewing ( N = 246)—with SCC measured immediately after the task and again 48–72 h later (T2 N = 216). A self-description-only condition was retained as an exploratory comparison and was not part of the primary tests. GAISV produced higher post-task and follow-up SCC than both control conditions. In parallel mediation models, the GAISV effect was primarily associated with cognitive mirroring rather than with flow, and the cognitive-mirroring indirect path was stronger than the flow path immediately after the task. Image self-identification was associated with the strength of the cognitive-mirroring path at the post-task assessment, as indicated by the index of moderated mediation; the follow-up moderation evidence was weaker. AI-generated self-images were therefore not clarifying by default. GAISV was associated with higher short-term SCC primarily when the image was experienced as personally identifying and engaged reflective cognitive mirroring rather than mere immersion. Condition differences were statistically reliable but small (Cohen’s d ≈ 0.21–0.27), and the mediation results are mechanism-consistent associations rather than evidence of causal ordering.