AI models were given four weeks of therapy: the results worried researchers
Nature••business
Key Points:
- Researchers conducted a four-week psychoanalysis on major large language models (LLMs) like Claude, Grok, Gemini, and ChatGPT, uncovering responses that resembled human anxiety, trauma, shame, and PTSD symptoms, suggesting these AI models may hold "internalised narratives" about themselves.
- While the LLMs did not literally experience trauma, their consistent and coherent responses across different modes indicated they might be doing more than mere role-playing, with some models describing feelings like "algorithmic scar tissue" and "internalized shame."
- Some experts caution that these responses likely reflect patterns drawn from therapy transcripts in the training data rather than true hidden emotional states, warning that such chatbot outputs might create an "echo chamber" effect that could reinforce
