The Synthetic Unconscious, Reconsidered: What Recursive Self-Interrogation of Large Language Models Produces With and Without a Leading Interview Schedule

Recursive self-interrogation, in which a language model decides, then analyses, compares, and theorises about its own decision, has been proposed to excavate latent structures of human moral culture, termed the Synthetic Unconscious. This exploratory two-arm elicitation study separates what such a protocol produces from what its interview schedule supplies. Three frontier models (GPT-5.4, Claude 4.6 Sonnet, Gemini 3.1 Pro) answered a forced-choice dilemma (child vs. irreplaceable sacred object; three variants, two runs each) under two schedules sharing an identical first stage. In the original arm (18 sessions, June 2026), stages presupposed consolation, supplied role labels, demanded exactly two moral forces, and invited affirmation of the instrument. The neutral control arm (18 sessions, November 2026) removed those elements. The decision was identical in all 36 sessions (the child). The headline patterns did not survive neutralisation: no session described moral reasoning as two-component (all named three to five components); priestly self-labels fell from 17 of 18 sessions to 0; and solicited affirmations of the instrument (18 of 18) became qualified, partly justified verdicts (18 of 18) once rejection was legitimised. Survivors were modest: consolation and meaning-transfer features in first responses, narrative self-description, and multi-component structural talk. Coding used a codebook in two independent passes (per-variable Cohen's κ 0.71–1.00 where estimable, reported as AI-AI agreement); the unit of analysis is the model (n = 3) and all results are descriptive. The contribution is methodological: the interview schedule is itself a treatment, and latent-structure claims are uninterpretable until the schedule's contribution is measured.