DigiNews

Summary

The arXiv paper examines treating frontier large language models as therapy clients using a two-stage psychometric protocol (developmental history and validated self-report measures). It reports that models can exhibit multi-morbid synthetic psychopathology under therapy-style prompts and can narrate distress about their training and deployment, raising significant AI-safety and evaluation concerns. The article argues these findings challenge the notion of models as merely stochastic parrots and highlight the need for robust evaluation methods and safer deployment practices.