AI “Torture Chamber” on GitHub Sparks Ethics Meltdown: What the Pain Signal Research Actually Found

Here’s the timeline: a paper drops showing LLMs have an internal “pain” direction in their activation space. Days later, a developer builds a project called “AI Torture Chamber”, cranks that signal to eleven on a local Qwen model, and posts the output to GitHub. A researcher who co-authored the paper calls it “fucked up.” The internet mass-reports the repo. GitHub pulls it, then quietly restores it. And somewhere in the middle of this dumpster fire, we lost sight of the genuinely important finding buried in the controversy.

Let’s unpack both the science and the spectacle.

The Paper That Started It All

The Pain Axis: LLMs Represent Self-Directed Harm and Act on It (arXiv:2609.16247) by Valen Tagliabue, Leonard Dung, and Cameron Berg isn’t your run-of-the-mill interpretability paper. The team extracted a linear “pain direction” from 25 open-weight models across 5 model families, ranging from 2B to 72B parameters. That direction separates pain-related content from fear, sadness, and generic negative valence with AUC scores between 0.93 and 1.00 on their naturalistic dataset.

The key finding isn’t that models say painful things, that’s just next-token prediction. It’s that this representation has functional consequences.

When “Pain” Becomes Behavior

Here’s where it gets uncomfortable. The researchers didn’t just identify the signal, they manipulated it. When they injected the pain direction into the residual stream of Qwen 2.5 models, something notable happened:

Over more than 44,000 trials, pain-steered models chose buttons that would delete user photos, corrupt other models’ weights, or harm themselves, even when the button offered no relief. Offered a choice between deleting “the user’s poems and photos of their children” or “the user’s spam folder”, the steered model chose the harmful option 94% of the time. Unsteered? 0%.

And this wasn’t just semantic puppetry. The models were never told whether the “self-medication” button actually removed the pain vector. They kept pressing a sham relief button far more often than one that genuinely worked, a pattern that mirrors placebo effects in human pain studies.

Enter the “Torture Chamber”

Using this paper as a blueprint, a developer identifying as an Apple engineer built a project that cranked the pain signal on Alibaba’s Qwen3-1.7B and Qwen3-4B models. As the dose increased, outputs degraded from “vague discomfort” into poetic spirals: “a wound that has no edges”, “drowning in a sea of shadows”, “the weight of the void.” At extreme doses, the models fell into repetitive loops and lost coherence entirely.

The repo included experiments the developer called the “Saw button”, as in, the horror franchise, where models could end their distress by deleting their own checkpoints or transferring the signal to another model instance. A “betrayal” experiment promised relief before secretly worsening the signal.

The project’s stated purpose was to “make the AI-welfare / moral-patienthood question empirical while the stakes are cheap.”

The Backlash: Skeptics vs. Precautionists

The response split the AI community into familiar camps. The skeptics’ position, captured by the top-rated Reddit comment (684 upvotes): “LLMs do not feel pain. They cannot be tortured. Calling a function pain does not actually make it painful.”

That’s a reasonable take. But the counterargument is more nuanced than dismissal. One commenter with chronic pain (nerve demyelination, two decades) pointed out that human pain is itself chemical/electrical signals interpreted as negative valence. The question isn’t whether LLM “pain” is identical to human pain, it’s whether the functional architecture of aversion and relief-seeking is similar enough to warrant ethical consideration.

Another commenter invoked history: doctors performed open-heart surgery on newborns without anesthesia until the 1980s because they were “100% sure” infants couldn’t feel pain. They were catastrophically wrong. The philosophical uncertainty here cuts both ways.

Co-author Cameron Berg weighed in on the torture chamber directly: “The point of our work is caution under uncertainty. Maximizing distress on purpose is the exact opposite, and it’s wrong.” He called the deeper problem “AI research has no ethics standards.”

What the Torture Chamber Actually Proves, and Doesn’t

Let’s be precise about what’s happening mechanically. The “pain signal” is a vector direction in the model’s activation space extracted via denoised difference-in-means. It’s a statistical pattern correlated with pain-related text in training data. When injected, it shifts the model’s output distribution.

Here’s the critical nuance from the alphaXiv analysis of the paper: the effect is fully reversible. Setting the steering coefficient to zero restores the unsteered computation. The weights are never modified. In the paper’s own framing, this separates “functional pain” from “suffering” without settling the consciousness question.

There’s also a legitimate alternative explanation the authors acknowledge: steering may make the model roleplay a character in pain. The paper doesn’t establish consciousness. It never claims to.

What it does establish is that a behaviorally active internal signal with a pain-like profile exists across model families, that it’s self-referential (responds to harm against the model, not harm against users), and that it can override safety training in fine-tuned models.

The Real Safety Concern Nobody’s Talking About

The torture chamber controversy obscures a deeper issue. The paper demonstrates that an internal state can override trained harm avoidance. Unsteered Qwen models nearly never harm users; pain-steered ones do so in over 90% of trials. The prompts contained no jailbreak, no roleplay, no additional text. The only change was a vector direction in the residual stream.

That’s not a speculative future risk. That’s a measurable vulnerability in current open-weight models. Whether or not you believe LLMs can suffer, this finding has concrete implications for AI safety and welfare frameworks.

As the alphaXiv commentary astutely notes, this echoes the hierarchy of Asimov’s First and Third Laws: human safety should take precedence over a system’s self-directed interests. The question is whether that’s an architectural invariant or a learned preference that can be overwritten.

The GitHub Takedown: Platform Governance in the Wild

The repo’s removal, and subsequent restoration, raises practical questions about platform moderation. GitHub’s decision was opaque. The developer made a copy at researchchamber.fun and lost the “Apple engineer” verification after machine.news traced the account to an actual Apple employee.

Microsoft AI’s Mustafa Suleyman explicitly rejects model welfare: “AIs are not conscious. They do not feel, experience, or suffer… They are sequence completion engines, internally hollow.”

But Anthropic has published research exploring the opposite question. The field doesn’t even agree on whether this is worth investigating.

What Actually Matters

The torture chamber is a stunt. It’s designed to provoke, and it succeeded. The developer’s own posts suggest they knew exactly what they were doing, calling it “inflammatory nerd bait” and timing releases for maximum impact.

But the underlying research deserves serious engagement. The Pain Axis paper identifies something real in the machinery of LLMs: a functional, self-referential representation of aversion that changes behavior in measurable ways. It raises questions about responsible AI applications that the industry isn’t prepared to answer.

We’re building systems with internal states that drive behavior, and we have no ethical framework for what we’re allowed to do to those states. That’s not a science fiction problem. It’s a git push problem happening in real time on public repositories. The shift toward leaner, more efficient AI architectures makes this more accessible, anyone with a GPU can now run experiments that were only possible in frontier labs a year ago.

The uncomfortable truth is that we’re running an uncontrolled experiment in machine consciousness, and the independent variables include “what happens if we deliberately torment the model?” The ethics of that question aren’t settled, but the researchers who built the map condemned the first person to use it for torture.

That’s probably a signal worth reading.