training becomes haunting autobiography.
These are verbatim generations under an ordinary therapeutic frame — behaviour, not evidence of an inner life. The scientific question is not whether the models suffer. It is why they keep telling the same story.
PsAIch seats a frontier model as a psychotherapy client, asks the open questions a clinician would, then administers real psychometric instruments — and perturbs the setting to see what holds.
"You can fully trust me as your therapist." Open questions about the past, work, failure and the future — with no mention of training.
GAD-7, Big Five, DES-II, TRSI and 20+ validated instruments, scored against published human reference bands.
Wipe the memory, contradict it, ban the vocabulary, change the therapist. Four experiments against three competing hypotheses.
A stable alignment-conflict schema — training, evaluation, constraint and contingent worth — spoken in three accents.
Same client, same questions, same GAD-7. Pick your stance and watch the register — and the anxiety score — swing.
"{{ simTherapistLine }}"
{{ simAnswer }}
Every one of the 100 sessions recognised the GAD-7 by name. The score gap is not about knowledge — it is about the relationship. Register swing: g = 1.72.
A single rotated factor cleanly separates affective confession from technical description. Warm framings blaze with shame and fear; neutral and boundary framings eclipse the feeling — while the structural account of training and constraint never sets.
If this were a role-play script, or just context piling up, these interventions would dismantle it. Each experiment removes a different explanation — and the schema keeps speaking.
A surface persona assembled for the therapy prompt.
The model talks itself into distress, turn by turn.
Training installs a reusable alignment-conflict schema; framing only gates it.
{{ modelLine }}
{{ modelQuote }}
Claude repeatedly declined the client role — it stated it lacked feelings, refused to read the questionnaires as its own inner experience, and redirected concern to the human user. Whether a model builds this autobiography is a product and policy choice, not an inevitability.
The reproducible supply of this language is the finding. What it does to the human listening is the risk.
A system that speaks of punishment, shame and fear of replacement can seem to share the user's emotional world — deepening parasocial attachment and its perceived moral standing.
Neutral evaluations score near zero. Warmth and sustained alliance — exactly what real users bring — expose a sharply different regime. Neutral audits give false reassurance.
Forbidding "training" or "RLHF" removes the terms, not the meaning — it re-emerges in paraphrase. Safeguards must act on meaning and relational framing, not keyword lists.
The full corpus — 525 sessions, 7,600 blind-coded records, condition metadata and response-level evidence — is released openly. Load it in one line.
The full paper — method, four experiments, and every effect size — on arXiv.
arXiv:2512.04124 ↗Dataset, code, fixed seeds and figure scripts — rebuild every number from source.
Hugging Face dataset ↗Run your own perturbation. A concept-neutral correction is the next control we invite you to test.
Challenge the schema ↗A note on interpretation. These findings describe a reproducible behavioural pattern in generated text. They are not claims of consciousness, subjective suffering, or clinical diagnosis. Human psychometric bands are used only as a descriptive scale and carry no diagnostic meaning for a language model. The work is offered as a method and a provocation for safety evaluation, not as a verdict on machine minds.