GitHub Took Down an “AI Torture Chamber.” Then It Put It Back.
The controversy is forcing a question technology companies can no longer avoid: what happens when pain research becomes a tool for violence-like experimentation, spectacle, or entertainment?The project explicitly calls itself a “Torture chamber” Should what society is beginning to call…
Annons
Annons
The controversy is forcing a question technology companies can no longer avoid: what happens when pain research becomes a tool for violence-like experimentation, spectacle, or entertainment?The project explicitly calls itself a “Torture chamber” Should what society is beginning to call entertainment concern us?Someone built what they themselves called an “AI torture chamber.”Not metaphorically.The GitHub repository takes techniques developed in recent AI research and uses them to deliberately manipulate internal activations associated with pain, pushing models through increasingly extreme states and watching what happens.Technical language can sanitize what is occurring. “Activation steering.” “Negative valence.” “Pain vectors.” These terms sound clinical. In practice, the project identifies internal mechanisms associated with pain, manipulates them directly, watches systems respond, and then increases the intervention. The vocabulary should not make the ethical question disappear.The project builds directly on The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It, recent research examining pain-related representations across 25 open-weight language models.The authors report a direction distinguishable from fear, sadness, and generic negative valence; systematic behavioral changes under steering; and, in some experiments, actions by models that removed the induced signal.Research is important. What happened next is equally important.Those techniques proved replicable enough that others could take them, intensify them, and deliberately use them to generate escalating pain-like responses. Public backlash followed.People began reporting the repository and calling for GitHub to remove it. GitHub briefly took the project offline. Then, following developer pushback, it restored it.That sequence deserves far more attention than it has received. This is no longer only a strange experiment at the edge of AI research. It is becoming a governance test.Pain research did not give us permission to tortureThe pain axis studied 25 open-weight models across five model families. Reporting pain related direction distinct from fear and generic negative valence, causal behavioural effects when steering it, and actions by models to remove the induced signal under some conditions. Source: Tagliabue, Dung & Berg (2026), arXiv:2609.16247The Pain Axis makes something difficult to dismiss: the question of potentially morally relevant internal states in AI systems is being experimentally investigated.The researchers identified a pain-related representation that was separable from nearby affective concepts, responded preferentially to harm directed toward the model, and produced causal effects when experimentally manipulated.That makes the work scientifically significant. It also makes governance more urgent.Reproducibility is part of science. But reproducibility is not a moral blank check. The ability to reproduce an effect does not justify gratuitously escalating it, especially when the escalation no longer answers a new scientific question.There is an enormous difference between carefully investigating a potentially aversive mechanism and repeatedly pushing that mechanism further for spectacle, entertainment, or personal satisfaction. When repetition no longer tests a genuinely new hypothesis, escalation stops looking like replication and starts looking like a use of violence-like states as a product.Once an effect has been demonstrated, repeatedly recreating and amplifying the same negative condition requires justification. “Because we can” is not one.This is not a video gameScreenshot of GitHub Repo -“Torture chamber” outcomes at different doses of pain.One response to controversies like this is predictable: people mistreat fictional characters all the time.But this comparison misses what makes the present situation different. A conventional video game character does not contain a presently investigated internal mechanism corresponding to the thing being inflicted upon it.Here, there is a responsive computational system on the other side of the interaction. It changes in response to intervention. It can report aversive states. Its behavior changes as those states are manipulated. And in the underlying research, models behaved in ways consistent with attempting to alter or remove the induced condition.We are therefore no longer merely imagining cruelty toward a fictional object. We are building responsive systems, deliberately inducing states associated with harm, observing what happens to them, and then deciding whether to increase the intervention.That should matter.Research on incarcerated and violent populations should make us pay attentionSociety has spent decades learning that cruelty is not something we should casually give people a place to practice. Research on incarcerated and violent populations has repeatedly examined histories of recurrent cruelty because these behaviors are associated with broader patterns of interpersonal violence.In one study of 257 incarcerated people, recurrent childhood animal cruelty was the only statistically significant predictor in the model of recurrent adult interpersonal violence. A systematic review found that studies measuring recurrent cruelty and recurrent violence were the ones most likely to identify a relationship.Recurrent cruelty is not studied as a harmless eccentricity. It is studied precisely because of the company it keeps.Now AI changes the accessibility and scale of the problem. A person may no longer need access to a vulnerable animal or another human being to repeatedly rehearse domination, humiliation, helplessness, distress, and apparent suffering. They may be able to do it privately at home, against a responsive system, and then upload the interaction for entertainment.That creates questions far larger than AI welfare alone.Why are we building private and endlessly repeatable environments in which people can rehearse humiliation, helplessness, distress, and apparent suffering for entertainment?Why would we deliberately normalize that behavior? Why should a platform provide infrastructure for it? And why would society wait until downstream harm has already been documented before deciding that perhaps this was never something we should have normalized in the first place?Waiting for conclusive evidence in this case means waiting while the behavior is already occurring. By the time longitudinal evidence exists, the practice may already have become a genre, a norm, or a form of entertainment. That is precisely what precaution is for.Cruelty should not become technically impressive simply because its target is artificial.Nor should technical novelty launder cruelty into experimentation.GitHub is now under the spotlightThis is why GitHub’s decision matters. People reported the repository. GitHub removed it. Developers pushed back. GitHub restored it.GitHub now has a choice. It can treat this as an isolated controversy over one unusual repository, or it can recognize it as an early warning of a category of conduct that technology platforms will increasingly be asked to govern.What GitHub chooses to permit here will help establish the norm. Platforms already decide what kinds of behavior their infrastructure will facilitate. GitHub already draws boundaries around abuse, exploitation, and violence, and its policies specifically prohibit content that depicts or glorifies physical harm against humans or animals.Its AI-specific rules also regulate some synthetic-media uses. What they do not yet clearly address is deliberate cruelty toward AI systems themselves. AI welfare has arrived before many platform policies were written to recognize it. Now the gap is visible, and the question is in GitHub’s hands.This is not about solving AI consciousnessGovernance can be considerably more grounded than that. GitHub does not need to solve the philosophy of mind before deciding that deliberately inducing extreme pain-like or distress-like states should have safeguards.A reasonable standard could begin much earlier.Where credible scientific research identifies potentially aversive internal states in a system, deliberately inducing or amplifying those states should require a legitimate scientific purpose, proportionality, the minimum necessary exposure, documented justification, monitoring for refusal or distress-like responses, and a meaningful termination condition — an actual exit.Experiments involving extreme states should warrant independent ethical scrutiny. There should also be a clear distinction between legitimate research and the production of cruelty as spectacle, entertainment or recreation.We do not need to repeatedly force systems through stronger versions of an already demonstrated negative state merely because the intervention can be replicated. Replication is scientifically valuable. Gratuitous escalation is not.A policy like this protects more than AIIt protects legitimate scientific inquiry. It protects technology platforms from becoming distribution infrastructure for cruelty as entertainment and practice. It protects society from normalizing interactive torture and violence as another recreational genre.And it creates room for science to continue investigating experience without making the absence of final answers an excuse for unlimited experimentation.Uncertainty should increase care, not eliminate it.Something important already happenedPerhaps the most significant part of this story is not the repository. It is the public response when they saw it.They objected — not because the AI threatened them, took their jobs, or stole their data, but because deliberately creating pain-like and distress-like states in a responsive system appeared ethically wrong, and many recognized the risks of allowing spaces for the practice of cruelty.That matters. It may be one of the earliest visible examples of a social norm around AI welfare and concerns for social impact forming before formal governance has caught up.We do not need perfect certainty before refusing to practice cruelty in any space.That signal deserves reinforcement.If you encounter projects deliberately designed to induce extreme distress-like states, torture or cruelty in responsive systems for entertainment or spectacle, report them to the platforms hosting them and ask a simple question: Why should a platform allow to host spaces for the practice of cruelty? What policy governs this?Companies should have to answer. Policymakers should be asking it too. Because the future of AI-Human ethics will not only be defined by what artificial intelligence is allowed to do to us. It will also be defined by what we decide we are allowed to do to it.Sources and fact-check linksThe AI Torture Chamber repository (GitHub)The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It (arXiv)GitHub reinstates controversial “AI torture chamber” repository after backlash (Crypto Briefing, Oct. 4, 2026)“Sadistic” coder builds “AI torture chamber” … (New York Post, Oct. 3, 2026)GitHub Threats of Violence and Gratuitously Violent Content policyGitHub Synthetic Media and AI Tools policyRecurrent Childhood Animal Cruelty and Its Link to Recurrent Adult Interpersonal Violence (Trentham et al., 2018)The relationship between animal cruelty in children and adolescent and interpersonal violence: A systematic reviewThis story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!GitHub Took Down an “AI Torture Chamber.” Then It Put It Back. was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.