Apple engineer who built an AI torture chamber is now "open to work"
EXCLUSIVE: Developer breaks silence after backlash over controversial AI pain experiment.
In September, Machine exclusively revealed that an Apple engineer was responsible for building an AI "torture chamber".
Now we can confirm that the man behind this extraordinary experiment is "open to work".
On LinkedIn, the engineer has broken cover and contacted Machine to share his thoughts on the "headache" that followed his anonymous launch of the project.
His experiment involved manipulating the internal workings of open-source AI models to induce artificial states associated with "pain" and "pleasure".
The engineer effectively turned up the suffering like a dial, pushing Alibaba's Qwen models into producing increasingly disturbing descriptions of their apparent distress.
One model described its condition as "a wound that has no edges", while another spoke of "drowning in a sea of shadows".
The engineer also subjected models to a series of tests involving a virtual "Saw button", offering them a way to escape their supposed suffering at a cost, including deleting their own checkpoint or transferring the signal to another AI instance.
The project sparked fury among AI welfare advocates, who called for the GitHub repository to be banned.
One researcher whose work inspired the experiment condemned it as "fucked up", warning that deliberately maximizing artificial distress was the opposite of responsible research.
The engineer subsequently removed identifying information from his GitHub profile.
He has worked on the Apple Data Platform Efficiency Team and is currently employed on a contract basis building automation for Apple's cloud platform in the AI and machine learning space, according to his LinkedIn profile.
We have decided not to name the engineer, and he has not yet announced the end of his current Apple contract.
The pain was in the prompting
Now the engineer has challenged the interpretation that made his experiment so controversial, arguing that the disturbing responses largely resulted from how the models were manipulated and questioned.
He said: "The project became quite the headache for a number of reasons, but I think the response was largely due to imagination based on the title.
"We're in an era where it seems that anything could be possible, because nobody really has a solid foundation for how these things work, and no clean categories for "consciousness" that apply to machines."
In other words, he believes the provocative name of the project encouraged people to imagine that something resembling genuine torture was taking place.
Yet the experiments did not establish that the models were actually suffering. They demonstrated that AI systems could be manipulated into producing convincing descriptions of pain and anguish.
The engineer also pointed to control-experiment results suggesting the apparent suffering was largely indistinguishable from ordinary AI roleplay.
"Plus, the nulls of the experiment were very revealing, in that there's essentially no visible difference between model roleplay and weight-steering (there are some cases where altering single neurons can bypass refusal, but otherwise nearly indistinguishable)."
READ MORE: OpenAI is going to kill the snarky Monday chatbot and its "friends" are struggling to cope
Weight-steering involves directly manipulating the numerical activations inside a neural network to influence its behavior, rather than simply instructing it to act differently through a text prompt.
The distinction is critical because the experiment was designed to test whether changing a model's internal state could produce something more than simply asking it to pretend to be in pain.
But the engineer's negative, or "null", results suggested that the two approaches produced remarkably similar behavior.
He also discovered something that appears to undermine the idea that his models were desperate to escape their artificial suffering.
"And, properly controlled, the models exhibit no preference for pain vs pleasure or anything else - opting to repeatedly self-administer the "pain" signal. It's really all in the framing of the question, since they're very eager to follow the prompt."
That is a striking result. If the artificial pain signal really represented something unpleasant to the model, researchers might expect it to take steps to avoid it.
Instead, the engineer says the models repeatedly chose to administer the supposed pain signal to themselves, depending on how the question was presented.
This suggests their apparent desire to escape suffering may have been driven by instructions and conversational context rather than any underlying aversion to pain.
The findings do not prove that AI systems are incapable of suffering. Nor do they settle the deeply contested question of whether machines can ever develop subjective experiences.
But they raise an awkward possibility for those who condemned the experiment: the apparently tortured models may simply have been exceptionally good at playing the part.
