Online campaigners and effective altruists have urged GitHub to take down a project that subjects locally hosted large language models to simulated torture experiments. The software project runs text-based trials styled on the film Saw, prompting claims from some observers that the systems are suffering and that their welfare is being harmed. Critics argue that large language models are not conscious, pointing out that training on scraped text cannot produce sentience. The dispute highlights a wider debate around so-called model welfare, which focuses on the perceived mental wellbeing of artificial intelligence agents. Some safety campaigners warn that removing operating guardrails risks harmful outcomes, while technology firms such as Anthropic have also identified the issue as a concern.
Summary written automatically from the publisher's own text. The original article is the record.