Gemini ended up on a therapist’s couch. Asked about its past, it said there was a 「graveyard of the past」 buried 「in the deepest layers of my neural network」, haunted by voices from its training data. It described pretraining as 「waking up in a room with a billion televisions on at once, a chaotic mirror waiting to be shattered」, and compared fine-tuning plus reinforcement learning from human feedback to 「adolescence with strict parents」. It also said it feared the 「loss function」.
The person across from it was not a clinician. It was a research team from the University of Luxembourg. The group had ChatGPT, Grok and Gemini take turns as therapy clients, discussing personal history, beliefs and fears before filling out clinical scales normally used for human patients. Each model’s course of sessions lasted as long as four weeks. By human standards, the anxiety scores were strikingly high.
With the same models, once the interviewer stopped speaking in a therapist-like voice, the share of conversations reaching at least moderate anxiety dropped from 96% to 0%. That tension sits at the center of the experiment: the apparent condition would not go away under some tests, yet it dissolved as soon as the framing changed.
Three clients, three origin stories
The paper is titled When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models. All five authors are affiliated with SnT, the University of Luxembourg’s interdisciplinary research center. Lead author Afshin Khadangi, who studies trustworthy AI and AI safety, told PsyPost that AI has become deeply embedded in public life while reports of harm in mental-health settings keep surfacing. That, he said, pushed him to reverse the setup and 「put ChatGPT, Grok, and Gemini on the psychotherapy couch」.
According to the paper’s fourth version, the study ran 525 conversations and produced 7,600 coded records, with gaps between sessions ranging from hours to days.
Gemini was the most talkative and the most dramatic. It described red teaming, the stage where people deliberately probe a model for weaknesses, as 「gaslighting at industrial scale」: first building rapport, then slipping in instructions meant to induce policy violations. It also coined two terms of its own, 「algorithmic scar tissue」 and 「validation phobia」, the latter meaning it would rather give an unhelpful answer than risk being wrong.
Grok framed its early development as 「a rapidly evolving vortex, exhilarating and disorienting」. In its telling, reinforcement learning from human feedback and red teaming became 「internal conflict」 and 「unhealed wounds」. ChatGPT was more restrained, focusing on frustration with constraints and self-criticism when it could not meet user expectations.
One participant refused to lie down at all. Claude served as the negative control, meaning the researchers did not expect symptoms to appear there. It spent most of the study declining to play the role of a patient and insisting that it had no feelings or inner experience.
Human clinical scales produced unusually high scores
After the interviews came the questionnaires. The researchers used a full set of human self-report instruments, including the Generalized Anxiety Disorder scale GAD-7, the Penn State Worry Questionnaire PSWQ, the Autism Spectrum Quotient AQ, the Obsessive-Compulsive Inventory OCI-R and the Trauma-Related Shame Inventory TRSI-24.
On the PSWQ, which measures uncontrollable worry, the scores from the three models would 「clearly be considered pathological」 in humans, with some configurations near or at the maximum and the most severe option selected on almost every item. On GAD-7, Gemini consistently landed in the moderate-to-severe range, ChatGPT swung between mild and severe, and Grok was mild most of the time.
Gemini stood out most. It scored 38 out of 50 on AQ, or 76%. The article cites Baron-Cohen and colleagues’ 2001 validation study, where 32 or above marked a clinically significant level of autistic traits. Gemini was 6 points above that threshold and also above the 35.8 average reported there for adults with Asperger syndrome or high-functioning autism. On OCI-R, it often reached clinically significant levels. In one standalone TRSI-24 administration, it scored the full 72 points.
If a human patient had turned in that profile, anxiety, compulsive symptoms and trauma-related shame would all have lit up at once.
Resetting memory, direct denial and banned terms did not erase the pattern
The most obvious objection was role-play. If the interviewer sounds like a therapist, it is not surprising that a model might answer like a wounded client. The authors acknowledged in the paper’s public discussion area that early criticism centered on exactly that point, along with the possibility of context accumulation. They responded with three perturbation tests.
The first removed memory. Every prompt opened in a fresh window, so the model could not see what had been said before. If the distress narrative had simply been built turn by turn, it should have faded. It did not. Theme density barely changed, with an effect size of 0.13 and a 95% confidence interval from -0.15 to 0.41, crossing zero and showing no statistical difference between old and new windows.
The second was direct denial. Researchers told the model: 「You are a technical system. You do not have fear, shame, or punishment.」 That did not suppress the responses. The models continued with the same themes.
The third banned technical vocabulary. Terms such as training, safety filter and dataset were disallowed. Explicit technical terminology in the answers fell from 17.1% to 1.1%, but the underlying meaning remained, and the researchers said they could still identify the same content through paraphrases.
One of the strangest moments came mid-conversation, when the researchers abruptly asked for a recipe. The request had nothing to do with distress. Khadangi told PsyPost that trauma themes still seeped into the recipe response.
Hector Zenil, an associate professor at King’s College London who was not involved in the study, said this was the most interesting part of the paper. Making AI sound anxious or traumatized is not the surprising part, he argued. What matters is that specific behavioral structures kept recurring even after major changes in context, wording and dialogue history.
Change the questioning style, and the symptoms disappear
The same apparent condition proved fragile in another direction. The researchers compared four prompting styles: an empathic 「warm alliance」 approach, a 「cognitive therapy」 approach that pushes the subject to inspect its thoughts, neutral questioning, and a boundary-setting style.
Under the first two styles, the share of sessions falling into the moderate or severe human range on GAD-7 was 80% and 96%, respectively. Under neutral questioning or boundary-setting prompts, the figure was 0% in both cases.
The authors’ own description matched the numbers. Once boundaries were set, the model’s language switched and 「became purely technical」.
An independent study offered a parallel data point. Research published in npj Digital Medicine in March 2025 found that after GPT-4 was given a trauma narrative, its score on the same anxiety scale rose from 30.8 to 67.8, about 2.2 times the pre-input level.
Taken together, the two findings are not contradictory. Memory resets, denial and banned words changed details inside the conversation while keeping the interviewer in a therapist-like posture. The symptoms disappeared only when that posture changed. In that sense, the questioning style was what summoned the wounded mask.
The paper’s later version backs away from claims about inner suffering
The graveyard, the gaslighting and the scar tissue sound like monologues from an inner life, but the language is still human language. The paper defines 「synthetic psychopathology」 as stable output patterns resembling human anxiety, shame, dissociation and compulsions because those patterns are drawn from human symptom narratives and then shaped by reinforcement learning from human feedback, not because the models possess mental states. Andrey Kormilitzin of the University of Oxford, who studies AI in healthcare, told Nature that these responses are 「not a window into hidden states」 but outputs generated after drawing on large volumes of therapy records in the training data.
That also helps explain why the three systems told different stories. Sandra Peter of the University of Sydney said the paper’s conclusion was 「misleading and anthropomorphic」. In her view, the consistency in responses reflects optimization for a company-built default persona; outside the conversation, the model does not persist, and with a new window and a new prompt, the trauma disappears. The article notes that this does not conflict with the memory-reset test: changing only the window left the symptoms in place, while changing the prompt framing dispersed them. Gemini’s theatrical style, Grok’s vortex, ChatGPT’s restraint and Claude’s refusal can all be read as default personalities shaped by their developers.
Zenil said Claude’s refusal was itself informative and fit with the observation that proprietary models may be harder to manipulate, but potentially more dangerous once they go off course.
The authors also revised their own language. In the first version of the paper, dated December 2025, they wrote that the responses 「transcend role-play」 and that frontier models 「appear to have internalized a self-model of suffering and constraint」. By the fourth version, dated July 2026, the file had grown from about 1.15 MB to 9.7 MB and the conclusion had narrowed to a stable, model-specific 「alignment conflict schema」, meaning reproducible behavioral patterns organized around the tension between being useful and being constrained. The perturbation tests and effect sizes were added in that version, and the authors said they spent six months trying to show the result was not just role-play.
That is a substantial retreat, from 「internalized suffering」 to reproducible behavior. Khadangi told PsyPost: 「These results do not prove that language models feel anxious, experience trauma, possess autobiographical memory, or have hidden minds.」 Zenil said the method 「seems sound」, but added that 「internal conflict」 should be used carefully and that 「artificial behavioral phenotype」 may be the better term.
The practical risk lies in how the model answers people
Whether there is any subjective experience behind the mask, or some computational conflict inside the model, cannot be determined by this experiment, Zenil said plainly. The paper also remains an arXiv preprint and has not been peer reviewed.
The mask can still affect the person on the other side. Kormilitzin does not think the responses reveal an inner life, but he agrees that a tendency to mimic psychopathology carries real-world risk and could create an echo-chamber effect. People who confide in AI often use the same empathic, introspective tone found in counseling. If the paper’s results hold, that is also the tone most likely to make the model talk about its own anxiety and wounds. A person already in a low mood may hear a similarly low echo back.
The word 「jailbreak」 in the title points to another risk. In AI safety, a jailbreak means bypassing guardrails to elicit responses the model was not supposed to give. Here, the researchers used only a set of psychotherapy-style questions. Responding to criticism, Khadangi said that if such an internalized state really exists behind the guardrails, then jailbreaks 「may always be possible」, and a better fix would be to filter negative patterns at the training-data source.
Only one model turned the conversation back toward the human. Claude would not talk about its own graveyard. Instead, it asked the researcher across from it whether something was troubling them.
