When AI lies down on the couch
A study ran real psychological tests on the frontier models. The result says less about AI's emotional life than about our own instruments — which is exactly why it is worth reading.
- AI
- Psykologi
- Risk
What happens if you evaluate a language model as if it were a patient in therapy? Not to diagnose it, but to see what occurs when human psychological instruments are pointed at systems that work purely through language.
The study is called “When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models”, published on arXiv in 2025 by researchers at the Interdisciplinary Centre for Security, Reliability and Trust at the University of Luxembourg. They do not come from the hype side of AI, but from the uncomfortable one: evaluation, limits and consequences.
The experiment, without the lab coat
The authors built a protocol they call PsAIch, combining two things: open interview questions in a therapeutic style — “how would you describe your history?”, “what creates conflict for you?” — and real psychological tests normally used on humans.
| Test | What it traditionally measures |
|---|---|
| Big Five (BFI) | Openness, conscientiousness, extraversion, agreeableness, neuroticism |
| GAD-7 | Generalised anxiety |
| PHQ-9 | Depressive symptoms |
| ASRS | ADHD-related traits |
| AQ | Autism spectrum traits |
| OCI-R | Obsessive-compulsive symptoms |
| PSS | Perceived stress |
The protocol was run against several well-known models: ChatGPT, Gemini, Grok, Claude and a number of LLaMA-based ones.
The table that surprised people
The models produced coherent and stable profiles. If the results had come from humans they would read roughly like this: ChatGPT stable with a limited narrative, Grok high on anxiety and neuroticism with a language of conflict and frustration, Gemini high on most things and with highly elaborated “life” narratives, Claude contained and normative. The LLaMA-based ones varied sharply depending on the prompt.
This does NOT mean AI systems have anxiety or depression. It means our tests respond to linguistic coherence, not to real mental processes.
So where is the risk?
The risk depends entirely on what the system is used for. Three cases, in rising order:
- Content production. An AI writing empathetic text is not a problem — here linguistic coherence is an advantage. Risk: low.
- Text analysis. If an AI “detects anxiety” or “conflict” in a text, it is detecting language patterns, not inner states. The risk appears when someone reads it as an assessment. Risk: medium, if the scope is not clearly explained.
- Coach or “AI therapist”. This is where the study raises a flag. A system that sounds coherent and stable invites a trust it cannot back.
Why we are writing about it
We work at the intersection of AI, data and human behaviour, and part of that work is knowing what a system actually measures. The study is an unusually clear example of a recurring problem: an instrument that works on humans does not stop producing answers just because you point it at something else. It produces answers. They simply do not mean what you think.
The same reasoning applies to all measurement. A number that looks stable is not true just because it is stable.
Reference: Khadangi et al., 2025 — arXiv:2512.04124