When Coaching Doesn’t Cash Out: Identity Salience and the Social Gate on AI-Taught Skills

Abstract: Large language models enable friction-free coaching, but a core question remains: when do AI-taught skills transfer from practice to real performance, and for whom? We examine an LLM-driven negotiation tutor that delivers diagnostic feedback by identifying dialogue-level negotiation errors with at least 90% accuracy. Across two studies, we uncover discontinuity between learning and payoff. In a practice context designed to minimize identity cueing, diagnostic feedback disproportionately improves performance for East Asian and White women, narrowing gaps. However, introducing an identity cue, telling participants that their counterpart is a White man, alters how learning converts into returns. Errors continue to decline, but outcome gains fragment along gender-by-ethnicity lines. These findings demonstrate that AI coaching can teach enacted skills, but whether those skills pay off depends on social context and identity salience. We conclude by outlining an identity-savvy design agenda for AI coaching as a sociotechnical infrastructure for equitable talent development.

Keywords: AI coaching; negotiation; diagnostic feedback; identity salience; backlash; intersectionality; learning transfer