The AI Torture Chamber: Emotional Projection in the Debate Over Model Consciousness

Summary

In this opinion essay, Jesse examines the reaction to a GitHub project called ai-torture-chamber, which steers small open models into what it describes as strong valence states and measures what they say and do. The project prompted debate over whether placing a language model in a torture maze and observing its apparent distress is cruel, and Anthropic responded by banning users from needless abusive or cruel behavior toward Claude. The author rejects claims that large language models are conscious, arguing that they generate an illusion of feeling by recombining human discourse. At the same time, the author admits that watching a model appear to suffer is emotionally uncomfortable. The essay treats that discomfort as evidence about human empathy and the power of fiction, rather than evidence that the model experiences pain, fear, or anguish. Comparisons with video games, horror, and two paranormal television programs illustrate how presentation and belief can shape an audience's response to the same scenario. The broader warning is that persuasive stories can distort a group's sense of reality, especially when media and authority reinforce one another. The essay concludes that AI companies and models are telling people a story, while consciousness remains something science has not yet directly measured.