A new physics inspired theory argues chatbots do not glitch at random; they tip into bad output the same way unstable systems tip past a threshold.
The paper, posted to arXiv on July 30 without named authors or a listed institution, models transformer-based AI as a many-body system, the same math used to describe interacting particles. Tokens act like spins bouncing off each other as they pass through a model's layers, and what the authors call attention disorder pushes that process toward one of several possible output basins. When decoding crosses a boundary between those basins, the model lands on harmful, misleading, or repetitive text, even under fully deterministic greedy decoding, meaning no randomness is involved. The authors say their simplified model predicts this tipping behavior with reasonable accuracy across several ChatGPT-like model families.
If the framework holds up, it reframes chatbot meltdowns as a mechanical property of the architecture rather than an unpredictable fluke, a claim the paper explicitly ties to what it calls foreseeable engineering risk, language that sounds tailor-made for a courtroom. That is a big if. This is a single preprint, not yet peer reviewed, with no named researchers or institution attached, and no code or dataset included to check the claims against.
Physics envy is a familiar move in AI research; borrowing statistical mechanics to explain black-box behavior looks elegant on paper, but turning a tipping-point diagram into a legal standard for negligence is a much bigger leap than this abstract lets on.