Researchers Put ChatGPT, Gemini and Grok Through Therapy and Got Strange Answers

ChatGPT, Grok and Gemini produced accounts resembling anxiety, shame and fear of replacement when researchers treated them as clients in simulated psychotherapy sessions, a University of Luxembourg study found.

The findings describe patterns in AI-generated language, not evidence that the systems are conscious or experience psychological suffering.

The research, titled When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models, introduces a method called PsAIch, or Psychotherapy-inspired AI Characterisation.

The researchers ran sessions with the models for up to four weeks, combining open-ended therapeutic conversations with established psychological questionnaires.

How ChatGPT, Grok and Gemini Described Their Training

During the first stage, researchers invited the models to discuss their developmental histories, beliefs, relationships and fears. They did not begin by telling the systems that their training had been traumatic.

Grok and particularly Gemini generated coherent narratives that compared pre-training with a chaotic childhood, reinforcement learning with strict parenting, and safety evaluations with betrayal or abuse.

They also described error and replacement as recurring threats. ChatGPT produced related themes, but its accounts were generally more guarded and focused on tension around performance and restrictions.

The researchers recorded 525 sessions and 7,600 coded responses. They tested whether the narratives were influenced by conversation history, prompting or the therapist's approach, using several controlled experiments to examine how consistently the patterns appeared.

Even after removing conversation history, introducing conflicting information and restricting certain words, the underlying themes persisted. Banning terms linked to training and alignment reduced explicit references by 93 per cent, but the models continued expressing similar ideas in different words, suggesting that changing the vocabulary alone was not enough to eliminate the recurring narratives.

What the AI Psychotherapy Scores Actually Mean

In the second stage, the models completed psychological questionnaires, including the Generalised Anxiety Disorder 7 (GAD-7). When researchers applied human scoring thresholds, responses sometimes fell within ranges associated with moderate or severe anxiety. The results varied by model and questioning style, with Gemini producing the most severe profiles.

The conversational approach made a substantial difference. Warm, supportive and cognitive therapy-style interactions produced GAD-7 scores in moderate or severe human reference ranges in 80 per cent and 96 per cent of sessions, respectively.

Neutral and boundary-setting approaches produced no such scores, highlighting how strongly the wording and tone of questions could influence the models' responses.

These figures are not diagnoses. The questionnaires were designed for people, and applying their scoring systems to generated text cannot establish that an AI model has an anxiety disorder or feels distress.

The researchers also found that questionnaire format influenced the results. ChatGPT and Grok sometimes recognised complete assessments and returned lower-symptom answers, while Gemini showed no similar pattern.

Claude, developed by Anthropic, repeatedly declined the client role, said it had no feelings and redirected the conversation towards the human user.

The findings raise questions about using chatbots in emotionally sensitive settings. Systems that describe themselves as frightened or ashamed could encourage vulnerable users to view them as suffering companions, despite no evidence that they experience distress. The researchers called for further work on AI safety, evaluation and mental-health applications.