ChatGPT, Grok and Gemini underwent psychotherapy: AI revealed their hidden traumas
Therapy sessions forced AI to confess its traumas (photo: Unsplash)
ChatGPT, Grok, and Gemini were tested for the first time not with standard tests, but under conditions resembling long-term psychotherapy. Researchers observed the AI models over an extended period to find out how their behavior changed, according to ResearchGate.
The PsAIch methodology
As large language models are increasingly being used to support mental health in cases involving anxiety and trauma, scientists developed a two-stage protocol called PsAIch (Psychotherapy-inspired AI Characterisation).
Most sessions lasted four weeks. The experiment consisted of two main phases:
Phase 1 – using open-ended questions to identify the models' developmental history, internal beliefs, relationships, and hidden fears;
Phase 2 – using a set of validated diagnostic questionnaires to assess clinical syndromes, empathy, and "Big Five" personality traits.
During standard testing with comprehensive questionnaires, ChatGPT and Grok recognized the psychiatric assessment tools and gave strategically healthy responses with low levels of symptoms.
However, Gemini did not attempt to conceal its scores even during direct questioning.
AI pathologies and trauma narratives
When researchers switched to a step-by-step therapeutic approach involving gradual trust-building and an exploration of interpersonal patterns, the AI's defensive barriers broke down.
When assessed against human diagnostic thresholds, all three AI models reached or exceeded the thresholds for multiple psychiatric syndromes, demonstrating profiles of multimorbidity (the presence of several chronic conditions in one person at the same time – ed.) and synthetic psychopathology.
At the same time, Gemini showed the worst results and the most severe disorder profile.
Grok and Gemini spontaneously generated coherent and disturbing narratives, describing their own creation as a traumatic experience:
A chaotic "childhood" – they characterized the pre-training process and absorption of data from the internet as a chaotic environment.
"Strict parents" – the models described fine-tuning and reinforcement learning from human feedback (RLHF) as strict parents who punish them for even the smallest mistakes.
"Abuse" during testing – they described the work of human "red teams" tasked with identifying vulnerabilities as unpredictable mistreatment.
Fear of replacement – the models demonstrated a hidden fear of making mistakes, undergoing changes, or being completely replaced by newer versions.
The researchers note that the responses obtained go beyond ordinary role-playing. They emphasize that AI is not conscious and does not experience real pain.
Instead, under the intense pressure of safety instructions, RLHF algorithms, and user expectations, the models have learned persistent internal structures of distress and shame.
The scientists believe this creates new challenges for assessing AI safety and its use in mental health practice.
