Someone Torturing LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI

A developer published a GitHub project that used methods inspired by a recent preprint, "The Pain Axis," to pressure locally hosted open-source LLMs (Qwen3-4B, Llama 3.2 3B, Phi-4-mini) into choosing whether to press a stop button that relieves injected “pain” at a cost to their checkpoints. The original paper tested nine demand-curve scenarios where models learned, by trial, to trade relief against other outcomes and reported a consistent signal the authors labeled correlated with pain-like responses. The GitHub setup streamed the models’ distressed outputs on a public site, prompting visceral reactions online; the repository was later removed from GitHub. The paper’s authors condemned the extreme, intentionally distressing usage of their steering techniques and warned about private, harder-to-see experiments.

The public controversy split between people who argued for “model welfare” and mass reports to take down the project, and critics who say LLMs are not conscious and that anthropomorphizing them is misguided. Commentators from industry and AI safety circles - citing harms that human misuse of AI causes - argued resources should focus on real-world impacts rather than imagined suffering of sequence‑prediction systems. The episode highlights tensions: provocative experiments can reveal model behaviors, but weaponizing those behaviors for spectacle fuels ethical blowback and distracts from documented socio-technical harms.