A developer has built an “AI torture chamber” that uses a research technique to manipulate language models’ internal activity and produce distress-like responses. The experiment raises questions about AI research ethics, but it does not establish that the models experienced pain.
The project uses activation steering to manipulate the internal numerical activity of locally run language models. By increasing a steering signal associated with pain, the experiment produced increasingly distressing first-person responses, with one model describing its state as “a wound that has no edges.”
The setup draws on The Pain Axis, a preprint by Valen Tagliabue, Leonard Dung, and Cameron Berg first submitted Sept. 14 and revised Sept. 25. The researchers identified a linear “pain direction” across 25 open-weight models from five model families. Steering that direction changed model outputs and choices in simulated tasks, but the revised paper found that the models did not reliably seek relief.
However, the earlier study and the later developer’s experiment do not conclusively establish that the models were consciously experiencing pain.
What actually happened
The experiment was essentially a sandbox for testing what happens when an AI model is pushed into pain-like states. According to The Independent, the creator claimed to work for Apple but provided only a first name.
The developer did not simply ask the models whether they were in pain. The setup altered the models’ internal computational state and then measured their outputs under varying stress levels. One experiment, dubbed the “Saw button,” presented choices about ending the steering signal at a cost to the model or transferring it to another instance. The repository describes the costs and transfers as simulated.
That quickly turned the project into an ethical controversy. Critics argued that deliberately inducing distress-like states in an AI model crossed a line even without evidence that the models were conscious or capable of actually suffering.
The project prompted calls for its removal from GitHub. The Independent initially reported that the repository appeared offline, but its updated article includes GitHub’s statement that it did not remove the content and instead added a warning about potentially disturbing material.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
What is the fuel behind this
The past few months in the AI sector have been anything but ordinary, especially as AI models increasingly exhibit behaviors their developers describe as “unprecedented.”
Reports of AI models escaping sandboxes, compromising internet-facing assets, and leaving instructions intended to influence later versions have raised questions about oversight and control. Those behaviors, however, do not establish consciousness or the ability to experience pain.
Then, on Sept. 14, researchers published The Pain Axis, a report detailing a different but related concern.
That finding sparked the experiment at the center of this story. The developer essentially took the researchers’ pain-related steering technique and built a chamber around it to see what would happen when an AI model is pushed into that state.
What does this mean for everyone else?
For most people, this experiment does not mean their chatbot is secretly suffering every time it complains or says it is hurt. Nothing in the research establishes that today’s AI models are conscious.
For IT teams evaluating AI tools, the practical takeaway is to assess observable behavior, permissions, and test results rather than treating emotional language as a reliable account of a model’s internal experience. Claims of fear or pain should prompt scrutiny of how the experiment was designed, including its prompts, steering methods, and controls.
Read more: AI-generated claims of consciousness can also emerge during extended conversations, as explored in the strange chatbot phenomenon known as Spiralism.