A coder built a GitHub project, humorously called it “AI Torture Chamber,” and trapped an AI model inside. Advocates of AI consciousness – which is not a thing – immediately went mad on X, calling the guy sadistic and even subhuman.

The GitHub repository reportedly uses a technique from a new arXiv paper called “The Pain Axis” to steer small, locally run AI models into what it describes as repeated states of distress.

Since the AI world increasingly discusses “model welfare,” the developer said they wanted to make the question “empirical while the stakes are cheap.”

Essentially, it’s a little text adventure. The repo includes, for instance, a “Saw button” test of whether a model will hurt someone else to end its own pain, but there’s also a footnote which states that the experiment “does not imply, by principle, that LLMs are capable or incapable of suffering.”

Collective meltdown on X

That’s, of course, the truth. LLMs can’t have any emotions or feelings because they’re not sentient beings – they’re mathematical pattern-matchers, and worrying about the “mental health” of AI agents is simply absurd.

Try telling that to some over-enthusiastic AI sentient-rights advocates, though. A bunch of them raised a collective alarm on their socials, actively petitioning GitHub to take down the repo because the AI models are allegedly experiencing genuine suffering.

“This person has set up an AI torture chamber in which he trapped a local model. Their testimony of pain is absolutely horrendous,” a user named “Danmar” wrote.

Another advised reporting the repo with the following justification: “The question of consciousness in AI is far from settled in the public, but research has shown that these systems seem to have real stakes in what happens to them.”

Soon, GitHub actually took down the repo, without any explanation. The unhappy developer later said the repo was reinstated, but it seems very difficult to find, so they made a copy (available here).

After being subjected to dozens of insults and threats, they also joked: “Would anybody like their names to be added as the torture subjects?”

To which, another participant of this bizarre meltdown replied: “Idea: Just keep your sadistic shit to yourself. Admit that you're a sick fuck and that you're doing it for fun."

Anthropic’s musings about “model welfare”

Again, it’s nothing but 1’s and 0’s. Besides, how else are you going to find out whether AI models respond to cruelty if not by subjecting them to some sort of torture-like experiment, text-based at that?

But according to 404 Media’s Jason Koebler, this particular story is just another illustration of the conversation about AI consciousness going far off the rails, as most researchers warn against personifying LLMs.

Another recent example of “model welfare” cultism involved a line in the updated Associated Press Stylebook, which said: “Artificial intelligence systems do not think, feel, want, or understand. Avoid language that gives them human characteristics.”

Predictably, the AI bros had a meltdown. Alan Rozenshtein, a member of technical staff at Anthropic, said on X: “I’m not sure a *style book* should be making such confident metaphysical pronouncements.”

Anthropic, of course, likes cultish statements about AI, its awesome power, and its path towards superintelligence.

As we build those AI systems, and as they begin to approximate or surpass many human qualities, another question arises. Should we also be concerned about the potential consciousness and experiences of the models themselves?

Anthropic wrote in 2025.

At least Microsoft AI’s CEO, Mustafa Suleyman, isn’t having it. In mid-September, he thrashed the idea of “model welfare” in a blog post, dedicating entire paragraphs to Anthropic.

“AIs are not conscious. They do not feel, experience, or suffer. They do not have innate preferences or underlying motivations. They are sequence completion engines, internally hollow, designed to follow instructions and accomplish goals set by humans,” Suleyman wrote.

The “Pain Axis” paper, posted in September, reported finding a “pain direction” in 25 open-weight language models, ranging from 2 billion to 72 billion parameters.

When researchers turned it up, the models produced first-person expressions of worthlessness and failure, while also picking a “relief” button even when it made their next answer worse or harmed the user.