Deep dive: can AI models feel pain — the ai-torture-chamber story

In late September 2026, a GitHub repo called ai-torture-chamber went viral for steering small open models (Qwen3, Llama 3.2, Phi-4-mini) into a "pain" state with an intensity slider, then was mass-reported and removed. The deep dive explains the technique from the paper "The Pain Axis: LLMs Represent Self-Directed Harm and Act on It" and why its authors called the project gratuitously cruel.

• The ai-torture-chamber repo was removed from GitHub after mass reports
• The technique comes from the paper "The Pain Axis: LLMs Represent Self-Directed Harm and Act on It"
• It locates a "pain" direction in the residual stream and amplifies it in Qwen3, Llama 3.2 and Phi-4-mini
• Co-author Cameron Berg called the project "gratuitously cruel" and stressed uncertainty about model consciousness