AI Tipping Dynamics: ChatGPT Failures as Foreseeable Risk

A new computer science paper on arXiv (arXiv:2607.25279) titled "Many-body Tipping Dynamics of ChatGPT-like AIs" investigates why ChatGPT-like AI systems unexpectedly produce undesirable content such as harmful, misleading, or repetitive outputs, even under deterministic greedy decoding. The authors demonstrate that these tippings arise from many-body interactions between tokens, modeled as spins crossing a finite-layer system. Tipping is described as a dynamical first passage process between competing output basins, with attention disorder controlling transport dynamics. A few-basin reduction yields a closed finite-layer threshold that shows good agreement across ChatGPT-like families. The findings suggest that such AI failures represent 'foreseeable engineering risk' rather than inherently unpredictable behavior, with implications for legal and societal assessments of AI harm. The paper is categorized under Computer Science > Artificial Intelligence.