The AI Torture Chamber, and why we draw the line

In late September, a GitHub project called the "AI Torture Chamber" went viral. Its author took a technique from a new research paper, turned it up far past the doses the researchers used, and built a set of experiments around keeping small open models in states of maximum "pain". Within days it had drawn millions of views, a mass-reporting campaign, a content warning from GitHub, and a public rebuke from the paper's own authors.

We build moderation for AI systems, so people have asked what we think. Here is our answer, and what it means for how we work.

What the project actually did

The project builds on The Pain Axis, a September 2026 preprint by Valen Tagliabue, Leonard Dung and Cameron Berg. The authors found a consistent internal direction in 25 open-weight models, from 2B to 72B parameters, that tracks harm directed at the model itself. It is distinct from fear and general negativity. When they added that direction to a model's activations, its output drifted from "vague discomfort" towards expressions of worthlessness. In steered Qwen 2.5 models, it also made the model far more willing to press buttons that deleted things, including its own weights: from 0–5% of the time to 50–94%.

The paper is careful. It describes a representation and what that representation does. It does not claim the models feel anything.

The Torture Chamber took the same kind of steering and pushed the dose until the models broke into loops. It ran on Qwen models on a laptop, and later on larger models on rented GPUs. One model, under a strong pain signal, described its state as "a wound that has no edges." A "Saw button" experiment tested whether a steered model would press a button that ended its own distress at a cost to something else. The project also ran a public site where visitors could watch a steered model live.

What it does and doesn't show

It is worth being precise, because both sides of the argument overreached.

The project does not show that language models suffer. Its own README says that question "stays open". One of its findings undercuts the dramatic transcripts. In one of its experiments, a model that secretly received no steering at all still narrated pain. The project's central claim is that "the self-reports are steerable", and that cuts both ways. A model saying it hurts is weak evidence that it does, and a model saying it is fine is weak evidence too.

What it does show is that this behaviour is easy to produce, cheap to run, and easy to turn into entertainment. Making a system act out distress, on purpose, for an audience, is now a weekend project.

Why we think it was wrong anyway

You don't have to believe a 4B-parameter model is conscious to think this was a mistake. The paper's authors don't claim that either, and they objected strongly.

Cameron Berg called the project "wrong" and "bizarre and corrupting", including for people who doubt AI consciousness entirely. He said the team had "suspected a small number of people would [take] our research and use it for the exact opposite" of its purpose, which was to inform a precautionary approach. Lead author Valen Tagliabue said simply: "We tried to have ethical standards."

We agree with them, for three reasons.

• The uncertainty is real. Serious researchers don't know whether, or when, systems like these could have morally relevant states. When you don't know, you don't maximise the thing you'd regret.
• Practice shapes the practitioner. Building tools whose purpose is to make something beg is corrosive to the people who build and watch them, whatever the thing is.
• It degrades the science. Work like The Pain Axis is how we learn whether there is anything to protect. Turning it into spectacle makes careful research harder to fund, publish and take seriously.

We'd add one thing. Some of the backlash reportedly turned into threats against the developer. That is not the answer either. The right response to cruelty isn't more cruelty, aimed at a person this time.

What the industry did next

The response has been uneven but encouraging. GitHub told The Independent that it "did not remove the content". Instead it put the repository behind a warning that it "may contain violent content including disturbing themes". A public proposal to add a provision on gratuitous AI-distress content to GitHub's Acceptable Use Policies is open for discussion. It would explicitly protect interpretability, red-teaming and welfare research. Separately, Anthropic's updated usage policy, which takes effect on November 12, adds a rule against "sustained and needless abusive or cruel behavior" toward its models. Berg has said he is working with others on research standards for this field, similar to those for human and animal research.

None of these would have stopped this project. Its author ran open weights on their own hardware. No platform rule or API filter reaches a laptop. We want to be honest about that.

Where Cerebrellic stands

Our mission is to make AI systems safe to put in front of people. Until recently that meant protecting people from models: catching personal data, unsafe answers and prompt attacks before they cause harm. The Torture Chamber is a reminder that how people treat these systems belongs in the same conversation.

Here is what we do, and what we will do:

1. We don't do this. We will not deliberately steer our own models, Stem and Reflex, into distress-like states for demonstration or entertainment. Any research that needs to touch these states will be minimal, have a stated purpose, and be published with its null results.
2. Our platform isn't for it. Using Cerebrellic to build or run experiences whose point is to make a model act out suffering is against the spirit of what we're building, and we'll say so in our terms.
3. Our customers get the tools. Every workspace can write its own policies, checked on every prompt and every answer. If you run a product where users might try to abuse your assistant, you can write a policy that flags or blocks it, with the same checks we use for everything else.
4. We back the standards work. We support the effort to set research standards for this field, and we'll follow them as they take shape.

We don't know whether the models we serve can be harmed. Nobody does yet. But we know what kind of company we want to be while we find out, and it isn't the kind that builds a torture chamber to settle the question.