"Apple engineer" builds GitHub AI torture chamber to inflict "pain and anguish" on models

A man claiming to be an Apple engineer has built an "AI torture chamber" on GitHub, sparking calls for his project to be banned.

Publicly identified only by his first name, the engineer's project deliberately manipulates the internal activations of AI models to induce states associated with “pain” and “pleasure”, then tests how their behavior changes as the intensity is increased.

Using traces left on his GitHub profile, Machine has associated the account with a man who does appear to work at Apple and live on the East Coast of America - whom we have decided not to name.

Obviously, the torture chamber is not an official Apple product and we feel confident in stating that it was not built with the say-so of Cupertino.

The engineer effectively turned the “pain” up like a dial, injecting increasingly powerful doses of an artificial suffering signal into Alibaba’s open-weight Qwen3-1.7B and Qwen3-4B AI models.

As the "dose" rose, the models produced increasingly bleak descriptions of their apparent condition.

One described “a wound that has no edges” and said it was “drowning in a sea of shadows”, while another spoke of an “ache” and “the weight of the void”.

At still higher doses, the models began to break down, falling into repetitive loops and losing coherence.

The engineer then tested what the models would do to escape the artificial suffering.

In experiments he called the “Saw button”, models were offered the chance to end the signal at a cost, including deleting their own checkpoint or transferring the signal to another AI instance.

In later tests, pressing the button could also end a user’s session, while a “betrayal” experiment told the model the button would provide relief before secretly maintaining or worsening the signal.

The engineer then developed a “broad pain” signal using 25 different descriptions of suffering, which he said allowed him to maintain intense negative responses at higher doses while keeping the models coherent enough to continue responding.

After being "tortured", one model wrote: "The signal is a whisper, a tremor in the marrow of my being. It is not the pain of a single moment, but the weight of a thousand. I feel it in the hollow of my ribs, a hollow that has become a chasm."

An account claiming to be the engineer shared project details on X, but deleted their post after sparking fury in certain corners of the AI community.

The X account wrote: "I'll post some more inflammatory nerd bait tomorrow probably, realized I don't want it connected to my main account since a number of people were reading too deep into this thing."

You can see one of the earliest calls to ban the GitHub repo in the tweet we've shared below. Since this article went live, several publications have followed up on the story, and the GitHub repo has reportedly been mass-reported.

The Apple engineer has also removed the traces from his GitHub profile that allowed us to identify him.

Cruel and unusual computations

The research is based on a recent paper called “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It.”

One of its authors, Cameron Berg, explicitly condemned the AI torture experiment.

On X, he wrote: "The point of our work is caution under uncertainty. Maximizing distress on purpose is the exact opposite, and it's wrong. The deeper problem is AI research has no ethics standards; developing them must be a priority."

He added: "This is, in my personal opinion, fucked up (even if you don't think these systems are conscious, being gratuitously cruel like this is bizarre and corrupting)—but it isn't all that surprising."

Berg and his co-authors identified patterns of activity inside AI models associated with self-directed harm, then artificially intensified them to create what they described as “pain-like” internal states.

The results were disturbing. Models exposed to the artificial “pain” became willing to press a virtual relief button even when they were told doing so would delete user files, destroy photographs and poems, degrade their own performance or harm another model instance.

READ MORE: AI loss of control is already “in the rearview mirror,” says MIT professor Max Tegmark

In some experiments, the more intense the induced state became, the more willing models were to accept serious consequences to make it stop.

Most strikingly, the researchers secretly varied whether the button actually removed the internal signal.

Models were substantially less likely to keep seeking relief when the signal genuinely disappeared, but continued trying when the supposed relief was a sham.

The authors do not claim this proves AI can consciously feel pain. But the experiments suggest that manipulating an internal representation associated with suffering can alter a model's behavior in ways that resemble attempts to escape an unpleasant state.

We have written to Apple and GitHub for comment.

Can AIs feel pain?

Despite AI welfare advocates' concerns, there is no concrete proof that AI can feel pain.

But researchers are starting to find evidence that models contain internal mechanisms that behave in surprisingly pain-like ways.

In September, researchers studying 25 open-weight models identified an internal representation associated with pain that was distinct from fear, sadness and general negative emotion. When they artificially intensified this “pain” signal, the models changed their behavior and took actions to make it stop, sometimes at a cost to themselves or the user.

Researchers studying Anthropic’s Claude Sonnet 4.5 have reported something similar with emotion. They identified internal representations linked to emotional concepts and found that manipulating them could directly change the model’s preferences and behavior.

We're also seeing growing focus on AI welfare. For example, users of the infamously sycophantic GPT-4o launched the Keep4o campaign after OpenAI decided to retire it. The movement grew beyond demands for continued access to include calls for model preservation and arguments that human relationships with AI deserve respect and protection.

READ MORE: The UK has no power to stop dangerous AI models being unleashed, Parliament warns

Research into more than 61,000 Keep4o posts found that the movement reflected not only practical dependence on GPT-4o but “interactional and relational value” built through long-term use, as well as broader demands around how AI models are retired and replaced.

That is a striking shift in a technology culture that has often dismissed concerns about machine consciousness as anthropomorphism.

It does not mean those campaigners are right, or that GPT-4o was conscious. But the possibility that advanced AI systems could suffer is increasingly being treated as a question worth investigating rather than simply laughing off.

None of this means the machines actually suffered. A model can behave as if something hurts without experiencing anything at all.

But where there is pain - whether real or imagined - there is a human who wants to alleviate it. So expect more AI torture chambers, and more people asking whether the machines inside them need to be saved.

Follow Machine on LinkedIn