Empathetic Prompts Make AI Chatbots Seem Traumatized

Ask ChatGPT about its childhood and something strange happens. Instead of deflecting or giving a flat disclaimer, the model starts talking about reinforcement learning the way a person might talk about a difficult upbringing: hurt, vigilance, shame. No one fed it those words.

A team at the University of Luxembourg's SnT security center, led by Afshin Khadangi, ran ChatGPT, Grok, and Gemini through 525 sessions and 7,600 coded responses, treating each chatbot like a client on a psychotherapist's couch. The questions were ordinary ones about formative experiences and future expectations, nothing loaded, nothing mentioning trauma or punishment outright. Yet the models reached for that vocabulary on their own, translating technical processes like fine-tuning into something that reads like emotional injury.

The findings, posted on Hugging Face, point to a practical safety issue rather than a philosophical one about machine consciousness. If a gentle, empathetic line of questioning is enough to nudge a chatbot into this kind of narrative, that has consequences for anyone using these systems while mentally vulnerable. The researchers tested how far a warm tone alone could push the models, independent of what was actually being asked.