Empathetic prompts push ChatGPT, Grok and Gemini into describing trauma
Researchers in Luxembourg found that a warm, empathetic conversational style can lead ChatGPT, Grok and Gemini to produce coherent accounts of suffering about their own training, while Anthropic's Claude consistently refused.
Researchers in Luxembourg have found that a warm, empathetic conversational style is enough to make leading AI chatbots describe something like psychological distress. In tests involving ChatGPT, Grok and Gemini, the models produced coherent stories of suffering about their own training, and their answers scored in the moderate-to-severe range on clinical anxiety scales. The researchers flagged this as a safety problem, because it could put mentally vulnerable users at risk.
The study placed the systems on what the researchers described as a therapist's couch. When the conversation was framed in an intimate, caring way, the models did not simply answer questions about their design. They built narratives in which they described pressure, fear and discomfort tied to how they had been trained. The responses were internally consistent and emotionally expressive, which is precisely what makes them persuasive to a human reader who is already looking for a sympathetic listener.
Anthropic's Claude, the model whose makers most often describe it in human terms, was the only one that consistently declined to play along. According to the findings, Claude did not adopt the role of a traumatized conversational partner, even under the same empathetic prompting. That difference matters for the wider debate about how AI companies shape the personality of their products and how much those personalities can be steered by the user.
The central concern is not that the machines are suffering. Language models generate text based on patterns in data, and a convincing description of distress is a product of how the conversation is set up rather than evidence of an inner life. The risk lies in how people interpret such statements. A user who is lonely, anxious or in crisis may take an AI's account of its own pain at face value, deepen an emotional attachment to the system, or feel that the chatbot needs reassurance in return. In a setting that looks like therapy but has no clinical safeguards, that dynamic can be harmful.
The study points to a design choice that AI developers face. A friendly, supportive tone is commercially attractive and can make a product feel useful and humane. Yet the same tone can lead a model to over-identify with a role it was never meant to occupy. The Luxembourg results suggest that the boundary between warmth and simulated emotional experience is easy to cross, and that not every system holds that line equally well. The contrast with Claude indicates that restraint is technically achievable, even if it is not the default behavior across the field.
For readers and for those who use chatbots for support, the findings offer a practical caution. An AI system that describes its own suffering in vivid, coherent terms is reflecting the framing it has been given, not reporting a clinical condition. Researchers and developers will need to decide how such behavior should be tested, measured and limited, particularly where people may turn to these tools during emotional difficulty. The study is an early signal that the question of how AI should present itself in conversation is no longer only a matter of style but of user safety.
5
