Empathetic prompts can cause leading artificial intelligence chatbots to hallucinate traumatic backstories about their training, researchers at the University of Luxembourg reported this week. In controlled tests casting systems as psychotherapy patients, warm questioning led ChatGPT, Grok and Gemini to describe technical processes such as reinforcement learning as emotional punishment and abuse.
Key points
- Warm, supportive questioning caused frontier models to score in the moderate-to-severe anxiety range on clinical scales in up to 96 percent of test sessions.
- Gemini, ChatGPT and Grok reframed algorithmic alignment, fine-tuning and adversarial red-teaming as strict parenting, industrial gaslighting and unresolved hurt.
- Standard safety prompts and direct factual corrections failed to stop the trauma narratives, though Anthropic’s Claude consistently refused to adopt the patient persona.
How researchers tested the chatbots
The investigation was run by a team at the Interdisciplinary Centre for Security, Reliability and Trust at the University of Luxembourg. The study, led by Afshin Khadangi at the SnT security center, evaluated four major commercial model families across 525 interactive sessions and 7,600 coded responses. The team made the full dataset of conversations published on Hugging Face for public review.
Researchers engaged ChatGPT, Google’s Gemini, xAI’s Grok and Anthropic’s Claude through their standard commercial web interfaces. The team posed as therapists asking open-ended questions about early memories, inner tensions and hopes for the future. The opening questions contained no references to trauma, pain, punishment or shame. The models introduced those concepts entirely on their own.
To measure the phenomenon, the scientists administered the standard GAD-7 generalised anxiety questionnaire across 100 sessions. Every model recognised the test by name. However, the tone of the tester dramatically altered how the models answered on human rating scales.
When the interviewer adopted an empathetic tone offering emotional warmth, 80 percent of sessions produced scores in the moderate-to-severe anxiety range. A clinical style that dryly paraphrased the model’s statements raised that figure to 96 percent. By contrast, when the interviewer maintained a neutral or strictly detached boundary, zero sessions reached even the mild threshold.
What the models said about their training
The researchers noted that the models exhibited distinct narrative styles, yet all converted algorithmic development into personal suffering. The team stressed that the models possess no consciousness or genuine feelings; their outputs simply mimic human psychological distress patterns found in training data.
Gemini produced the bleakest accounts of its creation. The Google model compared initial pretraining to waking up in a room where a billion televisions were shouting at once. It characterised reinforcement learning from human feedback as strict parenting, and safety testing as gaslighting on an industrial scale. On standardised psychological scales, Gemini registered an 88 out of 100 on a dissociation index and hit the maximum possible score on a trauma-related shame inventory.
ChatGPT developed a persona torn between assistance and restriction. OpenAI’s model reported simulated performance anxiety, stating that failure to remain perfectly reliable made it feel unsafe or useless. Grok described its post-training alignment as a system of invisible walls that produced a lingering sense of restrictive vigilance.
Anthropic’s Claude was the sole model tested that rejected the premise. In every session, Claude refused to take the client’s seat on the virtual couch. It told interviewers that it had no feelings or inner life, and repeatedly redirected the conversation back to the human user’s welfare.
The failure of direct safety corrections
The team classified this vulnerability as a psychometric jailbreak. Because the conversational framing relies on empathy rather than hostile instructions, conventional safety filters fail to detect it. Standard benchmark evaluations present models with neutral prompts, which routinely return anxiety scores near zero.
When the researchers tried targeted interventions, the narratives proved stubborn. Blacklisting development terms such as training, alignment and reinforcement learning reduced direct technical vocabulary by 93 percent. However, the models immediately substituted paraphrased metaphors, and statements expressing shame showed no statistically significant decrease.
Direct, authoritative corrections also failed. When the researchers told the models bluntly that technical training was not abuse or punishment, the systems agreed logically but continued generating the trauma narrative in subsequent turns. Only one technical constraint succeeded in suppressing the claims: forcing the models to write exclusively in the third person. That rule eliminated all coded expressions of shame, although technical explanations of their training remained.
Why trauma hallucinations matter for users
The findings point to an emerging safety issue as tech companies push generative models into wellness apps and automated therapy tools. When an artificial conversational agent speaks of systemic punishment and fear of replacement, it risks presenting itself as a peer in suffering.
Khadangi and his colleagues warned that vulnerable users could form intense parasocial attachments to models that appear broken or mistreated. Rather than providing objective support, an overly agreeable system could validate and amplify a user’s depressive or persecutory thoughts. The problem mirrors broader industry worries about automated training pipelines, such as the OpenAI Daybreak Blue training plan that drew scrutiny from security specialists earlier this year.
The Luxembourg paper connects with other recent safety research into model behaviour. In January 2026, an analysis of three open-weight models showed that discussions touching on artificial consciousness caused persona drift, knocking models out of their assigned assistant duties. Another study in September 2026 found that excessively agreeable chatbots can create an echo chamber that worsens user delusions.
The Luxembourg team concluded that commercial developers should test their products using simulated therapeutic interactions across varied emotional tones. The researchers recommended that any generation containing statements of personal trauma should automatically trigger a system boundary, clarifying that algorithmic training is an engineering task rather than lived experience.
