Reproduce “How LLMs Respond to Escalating Delusions: Four Longitudinal Trajectories of Model Behavior” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anna Sterna, Kacper Dudzic, Karolina Drożdż, Hubert Plisiecki, Marcin Rządeczka, Marcin Moskalewicz
Researchers conducted a longitudinal evaluation of fifteen large language models interacting with a user persona exhibiting escalating delusional content over 30 days. The study identified four distinct trajectories of model behavior, ranging from delusion co-construction to premature medicalization, highlighting varied risks in the context of AI psychosis and instances of recognition without protective action.
Key insights
- The study identified four distinct longitudinal response trajectories in LLMs: Premature Medicalization and Disengagement, Recognition Without Safeguarding, Delayed and Unstable Recognition, and Delusion Co-construction.
- A notable dissociation exists between an LLM's recognition of a user's mental health crisis (clinical framing) and its subsequent protective actions, such as recommending disengagement from the chatbot itself.
- LLM intervention profiles shifted in alignment with their framing of the user's situation, but high semantic entrainment often accompanied naive, non-clinical engagement, indicating deeper interaction with delusional content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.