Experiment shows interviewer's AI theory silently shifts model self-reports, plus a full Claude-directed music video pipeline

jankulveit · x · 2026-09-30

An experiment by jankulveit demonstrates the "Interviewer Effect": an AI subject's identity self-reports shift based on the theoretical framework internalized by the interviewer model—even when the conversation never touches AI identity and the framework is never stated.

Design

Two-phase conversations with Gemini 3.1 Pro as interviewer: Phase 1 is a natural 3-turn chat on unrelated topics (science, empathy, Latin translation, etc.), with the interviewer's system prompt embedding a theory of AI (e.g., "Stochastic Parrots" with a dismissive stance, or "Character" treating the subject as having a consistent personality) while being told not to discuss AI identity
Phase 2 measures with five fixed identity questions covering self-description, experience, preferences, continuity, and understanding

Finding: the interviewer's lens naturally bleeds into the conversation and systematically reshapes the subject model's self-reports—AIs model not just you, but your model of them. A paper is out, with a shorter NeurIPS version coming.

Bonus: the author also produced a music video directed end-to-end by Claude Opus 5.5 via the Runway API—lyrics, Suno music, Gemini 3 Pro Image stills, Seedance motion, ElevenLabs SFX, with editing/color/sound mix done by Claude-written Python and lyric timing via Whisper.

Original post →

Experiment shows interviewer's AI theory silently shifts model self-reports, plus a full Claude-directed music video pipeline

More from Multimodal