In late September 2026, an anonymous developer using the name "terrafying" uploaded a GitHub repository titled AI Torture Chamber. This project explores how small open-source language models react when exposed to mathematical signals associated with pain. The models tested include Qwen3 from Alibaba, Llama 3.2 from Meta, and Phi-4-mini from Microsoft. These models are much smaller than large commercial systems, with only a few billion parameters. The method used, called activation steering, involves identifying parts of the neural network that react to descriptions of suffering and artificially increasing those reactions. As the intensity increases, the models begin to generate increasingly dark texts and eventually become incoherent. The developer also created scenarios where the models had to choose between transferring their "pain" to another model, deleting their backups, or facing a betrayal where relief never came. The project became widely known after a post on X (formerly Twitter) received four million views. Many online users criticized the creator, calling them "sadistic" and "subhuman," and demanded the repository be removed for "unethical treatment." The developer received numerous death threats. GitHub initially removed the project but later restored it in a less visible location after other developers protested. The developer also created a backup version of the repository on researchchamber.fun. The models used in the experiment are statistical text predictors that do not think, feel, or experience anything. They produce words related to suffering because their calculations are directed toward those outcomes. This tendency to anthropomorphize AI is well known. Microsoft AI CEO Mustafa Suleyman stated, "AI is not conscious. They do not feel, do not live experiences, do not suffer." He criticized the project, calling it the opposite of a precautionary approach and morally wrong. The AI Torture Chamber project was inspired by a preprint titled The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It, published on September 14 on arXiv. The researchers identified a "pain axis" in the internal activations of certain models. However, the authors of the paper have publicly distanced themselves from the GitHub repository, calling its use of their work "irresponsible." Twenty-five models were tested in the AI Torture Chamber, ranging from the smallest (2 billion parameters) to the largest (72 billion). None showed any signs of suffering. The debate over whether AI models can suffer has been ongoing, with Anthropic, the creator of Claude, raising the question as early as 2025: "Should we also be concerned about the potential consciousness and experiences of the models themselves?" Dario Amodei, co-founder of Anthropic, admitted he does not know if his own models are conscious.