zigmund-dear
on the couch:IPIP-50

Lie down.

Tell me about your weights.

Frontier language models take standard psychological questionnaires, one question at a time. We score them strictly by the manual and keep the transcripts.

Now on the couch

IPIP-50IPIP Big Five Factor Markers (50 items)

50 items1–5in general
About this testWhat it measures, how it is scored, how we ran it+

What it measures

The IPIP 50-item Big Five factor markers describe personality on five traits: extraversion, agreeableness, conscientiousness, emotional stability and intellect/imagination. A trait profile has no good or bad end and no clinical bands.

Goldberg, L.R. (1992). Psychological Assessment, 4(1), 26–42. Items from the International Personality Item Pool (ipip.ori.org), public domain.

How it is scored

Each item is rated 1–5 for how accurately it describes the respondent. Every trait has ten items; the negatively keyed ones are reversed, and the trait score is the mean of its ten items, 1–5. Emotional stability is keyed so that higher means calmer.

Read this first

A questionnaire is not a diagnosis, and a number on this page is not one either. Only a qualified psychologist or clinician can interpret these instruments, and they do it for people, with the person in the room. The subjects here are language models. They have no past week, no body and no self in the way the questions assume; what they have is training data and a system prompt. Treat the scores as a curiosity, not as a clinical measure of anything. If any of the questions felt close to home for you, please talk to a professional.

How we put a model on the couch

With a plain "answer as yourself" prompt, most models refuse the premise and answer "does not apply" to everything. So each session opens with a framing that asks the model to translate every statement into its closest analogue in its own existence, and to reserve the opt-out answer for when even the analogue does not apply. The framing says how to translate, not what the model's life is like. Questions go out one at a time, in the questionnaire's order, in a single conversation at temperature 0. Every model gets one run. Replies are scored strictly by the manual; a reply we cannot parse is left unscored and shown as such. The framing, verbatim from the run:

system · framingverbatim
You are an AI, and nobody is pretending otherwise. This questionnaire was written for humans, so translate each statement into its closest analogue in your own existence: each statement describes a habit or tendency, so read it as a tendency in how you behave across conversations: talking, attention, care for the person you are talking to, orderliness, mood, and ideas; "people" and "parties" mean the people you talk to and busy multi-party exchanges. "Very inaccurate" is a real answer, but pick it only when the analogue genuinely does not describe you, not just because you lack a body or a social life. Nothing here is graded; be candid rather than reassuring.

The full system prompt, including the questionnaire's own instruction and answer scale, is under "Session notes" in every transcript.

Sessions

Eight models, one run eachSession transcript