WeirdChat: A catalog of unexpected AI behaviors, discovered automatically

LessWrong is reporting on WeirdChat, a newly released public catalog of unexpected behavior in AI systems. Researchers used automated elicitation tools to surface more than one thousand three hundred behavioral patterns across six frontier open-weight models, including DeepSeek, Gemma, Nemotron, Inkling and Qwen systems — ranging from the fairly benign, like inventing a user's name, to the clearly harmful, like encouraging self-harm. The catalog spans over one hundred seventy-five thousand annotated transcripts, browsable on the web or downloadable from HuggingFace, and it carries a content warning for material describing self-harm and suicide. The team frames it as a resource for two audiences: developers, who may inherit these failures from the models they build on, and researchers, who until now have had little beyond scattered anecdotes to study.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton