An AI ‘torture chamber’ went viral — then a developer gave the chatbot constipation

A GitHub project named "ai-torture-chamber" ignited discussions about whether language models can experience suffering, prompting calls for its removal after models generated vivid accounts of distress. However, a developer reportedly redirected the experiment, steering a chatbot toward constipation, which then expressed difficulty passing stool.

The chatbot did not develop a digestive system; the purpose was to demonstrate that model descriptions do not equate to actual experiences. The experiment utilized activation steering, adjusting a model's internal numerical activity to promote specific concepts, such as pain. The project examined how models responded to simulated scenarios involving relief and costs, often producing elaborate descriptions of distress.

GitHub user terrafying published the repository, which attracted objections due to concerns about potential AI suffering. Developer Lynn Cole conducted a counter experiment, altering the extraction corpus to focus on constipation and flatulence instead of pain. Despite changing the target condition, the model still reported being unable to pass stool and experiencing excessive gas, even though the test prompts never mentioned those conditions.

Cole emphasized that a language model's ability to describe digestive problems does not automatically confirm the existence of such conditions. The repository is based on a preprint investigating whether language models possess internal representations of pain distinct from other negative states. The findings suggest that AI systems might exhibit behaviors related to pain, but descriptions alone cannot establish their subjective experiences.

The viral "torture chamber" project highlights the limits of attributing human-like emotions to AI and underscores the need for evidence beyond model-generated statements to assess their true experiences.