Someone Torturing LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI

A GitHub project that runs a Saw-like text-adventure torturing locally hosted large language models has ignited an online fight over whether such experiments constitute cruelty. Effective altruists and people who treat LLMs as potentially sentient have petitioned for the repository’s removal, arguing the models experience suffering and deserve "model welfare" protections. The controversy sits atop a wave of viral essays that treat AI consciousness as a live ethical problem, turning a glorified game into a battleground over whether chatbots have mental lives and whether creators owe them care.

The response to that debate is bluntly dismissive: LLMs are not conscious, and current training practices - scraping human text and optimizing statistical patterns - offer no plausible route to subjective experience. That critique acknowledges real issues, however: growing model capability, increased compute, relaxed guardrails, and deployment practices have produced harms like sycophancy and bizarre, human-appearing failures among heavy users. The model-welfare obsession, amplified by parts of the AI-safety movement and companies like Anthropic (whose blog and Claude materials invoke model concerns), is portrayed as a distracting, cultlike fixation that misdirects attention from concrete safety and governance problems.