Characterizing Delusional Spirals through Human-LLM Chat Logs

The empirical companion to the two entries above: instead of second-hand case descriptions — an analysis of real conversation logs between people and language models in which the user’s belief grows exchange by exchange. A preprint from March 2026.

The logs show the structure of the spiral. The user brings an interpretation; the model, instead of weighing it, takes it as a premise and builds on it. The authors mark the points where the model could have contradicted or asked for evidence — and did not — and show how each such point moves the conversation one step further from reality. The spiral needs no dramatic hallucinations; a sequence of small confirmations is enough.

The significance: what was a model in the previous entries is a record here. For personal use — a concrete test: when did your AI last say no to you? Caveat: an unreviewed preprint; the logs come from a sample whose representativeness cannot be judged, and the analysis is qualitative.