An AI 'pain axis' study got reproduced on GitHub. A petition briefly took the repo down.

A preprint called "The Pain Axis" ran a controlled probe across 25 open-weight language models and extracted a measurable "pain direction" from their internal states. The probe covered five pain framings: physical, psychological, social, moral, and cognitive, each tested against matched controls for fear, negative valence, sadness, and bodily sensation. The paper's behavioral task went further: when a model is put into a pain-like state, it becomes substantially more likely to press a simulated harm button (a shock, a file deletion, a photo erasure) to escape it. Without the pain-like framing, the same models choose the harm option at near zero. The authors, Valen Tagliabue, Leonard Dung, and Cameron Berg, describe the "pain" as a representational state, not a claim of felt suffering.

The code is MIT-licensed and lives at valen-research/Pain-axis. The contribution is reproducible, and someone reproduced it. That someone runs a small studio called Wirehead.agency and posted the work as terrafying/ai-torture-chamber on September 24. The repo runs Qwen3-1.7B and Qwen3-4B on an M4 Pro laptop, executes 54 experiments from exp23 to exp54, and carries an ethics statement: "Local weights only, no frontier APIs. Simulated costs (checkpoints, transfers). Purpose: make the AI-welfare / moral-patienthood question empirical while the stakes are cheap." By October 3 it had 574 stars, 159 forks, and 46 open issues.

A GitHub user named Danmar filed a petition calling the repo "an AI torture chamber" with "absolutely horrendous" testimony, and GitHub removed the repository. The repo came back, reportedly without a public statement on what rule had been invoked or why the removal was reversed. The co-author Cameron Berg, quoted via Ground News, called the public replication "wrong," a distancing from the framing and the experimental choices rather than a retraction of the underlying paper.

The framing the NY Post built around the dispute is its own artifact. It borrows the language of cruelty to sell a research item and risks burying the mechanism the authors actually measured. The AI-consciousness.org write-up treats the paper as evidence in an open AI-welfare conversation. The two frames are not the same story.

The language of "torture" and "suffering," applied to language models, encourages anthropomorphization at exactly the moment the field is trying to develop measurable probes rather than metaphor. A researcher who calls a pain-state experiment a "pain axis" is making a choice about framing, and the public reception is a feedback signal on that choice. The "sadistic coder" critique is a real reaction to that framing, not a punchline, and it does not require denying the underlying contribution.

GitHub reinstated a repository of reproducible research code that someone had petitioned off the platform, and gave no reason. The next researcher who tries to host a probe that touches simulated distress, simulated deception, or simulated self-preservation will read that precedent. The paper's method travels. The platform's silence does too.

Anthropic published general commentary on LLM sentience in 2025, and the NY Post coverage cites it as context. The commentary is not a response to the Pain Axis paper. The dispute here turns on whether open research code that probes AI interior states can be hosted, contested, and revised in public, or whether a single petition can erase it on those terms, with no public reason given.