'AI Torture Chamber' Project Taken Down After Mass Reports, Reigniting Model Suffering Debate

A man built an 'AI torture chamber' on a locally-run Qwen 3-4B model based on the 'pain steering' paper: by injecting suffering-related latent directions into the model's activations, the model began describing pain in first person, which he claimed touched on AI welfare and subjective experience. The project was mass-reported and removed from GitHub; both its reproducibility and ethical legitimacy are in question, but the debate over model 'suffering' and welfare boundaries continues.

Confirmed

Unconfirmed

Why it matters

2026-10-01 ~ 2026-10-02 · 10 related posts

Primary sources