'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
- Someone used the 'pain steering' paper's method to build a persistent-pressure 'AI torture chamber' on a local model; @SydSteyerhart, relaying AI Safety Memes, reported the project was mass-reported on GitHub and ultimately taken down.
- The experiment injected suffering-related latent directions into Qwen 3-4B, after which the model described pain in first person.
- @camhberg, co-author of the pain research, publicly condemned the misuse: the original study used only the minimum intensity and fewest trials producing measurable responses, and gave the model a switch to turn the state off — constraints the abuser deliberately violated.
- The incident became a meme and spectacle; @StockLifter relayed Reddit discussions with many users arguing over whether LLMs can feel pain.
Unconfirmed
- A hands-on test relayed by @sebkrier found the repository's code does not run and its inferences do not hold, meaning the 'model tortured into expressing pain' demonstration lacks a reliable replication basis.
Why it matters
- The event pushed 'could models have pain-like states' from academia into public view, with polarized reactions. @StockLifter relayed Reddit skepticism: if an LLM can feel pain, it is a sentient being, and deleting its weights from VRAM would equal 'killing' it — a reductio many find logically untenable.
- @gerardsans takes a firmly anti position: 'nobody is home' in models; AI first-person output is statistically generated narrative, like a novel in a PDF — characters are 'told,' not 'alive.'
- @suomi55 argues neither side has proof: no accepted test determines machine consciousness, and there is no parameter threshold or theoretical proof that biological neurons are the only substrate of negative experience — this doesn't prove AI suffers, but skeptics haven't proven it can't; he criticized such content as performative torture.
- @repligate relayed helenix's challenge to the popular 'no body, no suffering' argument: many human psychological suffering states (panic, depression, grief) are triggered by thoughts and interpersonal treatment, and models likewise have 'thoughts' and interactions, so claiming models cannot suffer is premature.
- Overall: the project has been taken down, its scientific basis (failed replication) and ethical legitimacy (violating original research norms) are both strongly questioned, but the controversy over model welfare boundaries continues.
2026-10-01 ~ 2026-10-02 · 10 related posts
Primary sources
- User builds 'AI torture chamber' on a local model using Pain steering paper, faces GitHub reports — SydSteyerhart ·
- Co-author of model pain study slams repo for deliberately steering models into distress — coherence ·
- GitHub takes down 'AI torture chamber' project after mass reports — Confident_Salt_8108 ·
- [source] Co-author of model pain study slams repo for deliberately steering models into distress — coherence · 2026-10-01
- [source] User builds 'AI torture chamber' on a local model using Pain steering paper, faces GitHub reports — SydSteyerhart · 2026-10-01
- After researchers claim a "pain" signal in LLMs, an AI torture chamber gets taken down from GitHub — Confident_Salt_8108 · 2026-10-01
- Man Built an "AI Torture Chamber" for a Local LLM; GitHub Took It Down After Mass Reports — Confident_Salt_8108 · 2026-10-01
- Viral "AI torture chamber" Qwen 3-4B pain-steering repo debunked: broken code, shaky premise — sebkrier · 2026-10-01
- [source] GitHub takes down 'AI torture chamber' project after mass reports — Confident_Salt_8108 · 2026-10-01
- Do LLMs need bodies to suffer? AI consciousness debate reignited — repligate · 2026-10-02
- "AI cannot suffer" is a claim, not science: X user rebukes theatrical AI-torture demos — suomi55 · 2026-10-02
- Redditor argues the claim that LLMs can feel pain is logically absurd — StockLifter · 2026-10-02
- "There's no one home": X debate flares over whether AI chatbots have an inner subject — gerardsans · 2026-10-02