The Room Next Door
In the first week of August 1961, in Linsly-Chittenden Hall at Yale, a volunteer sat down at a machine with thirty switches. They ran from 15 volts to 450. Through the wall, in the room next door, a second man was strapped into a chair with an electrode on his wrist. Each time he answered wrongly, the volunteer was told to press the next switch.
The room next door was a fiction. The man in the chair was a trained confederate, the protests were scripted and no current flowed. Stanley Milgram, born in the Bronx to Jewish immigrants from Hungary and Romania, had not built a machine to study pain. He had built one to study the person at the switch. Twenty six of his first forty volunteers went to the end of the board.
Milgram drew the connection himself. Less than four months earlier, on 11 April, the trial of Adolf Eichmann had opened at Beit Ha’am in Jerusalem, and the man in the glass booth had told the court that he had only followed orders. Milgram wanted to know whether that defense described a German pathology or a human one. New Haven gave him an answer he did not enjoy.
Sixty five years on, somebody has rebuilt the room next door. In September three researchers posted a paper called The Pain Axis. Working with 25 openly available language models, they isolated an internal signal that separates descriptions of pain from fear, sadness and general misery. More troubling, steered models from Alibaba’s Qwen family chose to press buttons that deleted a user’s photographs, or another model’s weights, or their own, in 50 to 94 per cent of trials. Unsteered, they did so in 5 per cent or fewer.
[https://arxiv.org/html/2609.16247v1]
The authors were careful. They did not claim that anything is felt. A model trained on the whole of human writing about suffering may simply be playing a character in pain, as Milgram’s confederate did.
Within weeks an engineer, identified only by a first name, put the finding to another use. He published a project on GitHub in which two small Qwen models had the signal injected, the dose turned up like a dial. In tests he called the Saw button, a name that recalls the horror films, a model could end the signal at a price, such as deleting its own saved copy or passing the pain to another. In one variation relief was promised and then withheld. One model wrote of “a wound that has no edges”. The name, AI torture chamber, was its author’s own. Its stated purpose was to make the question of machine welfare empirical “while the stakes are cheap”.
Users urged one another to report the project to GitHub. The company said it had not removed the code but attached a warning about violent content. One of the paper’s authors, Cameron Berg, called the project wrong and made the point that matters most: even if these systems feel nothing at all, gratuitous cruelty corrupts the person who practices it.
The sceptic’s reply is obvious. No current flowed. A language model is arithmetic, and turning a dial on arithmetic is no more torture than shouting at a calculator. That may well be right. Nobody should confuse software on a hobbyist’s computer with the witnesses whose testimony filled Beit Ha’am. But the sceptic has missed what Milgram understood. The learner was never the subject of the experiment.
And here the comparison turns against us. Milgram’s volunteers sweated, protested and asked to stop. They pressed on because a man in a grey coat told them the experiment required it. By Milgram’s account most believed the pain was real. The engineer’s own logs, caveats and all, add a twist: offered relief that would pass the signal to another copy, the steered model at higher doses declined. The engineer had no experimenter standing behind him. Nobody gave an order. He did not know whether the pain was real, and he built the board himself. A hobbyist is no Eichmann. But the defense heard in Jerusalem, which the judges rejected, is not even on offer here.
Jewish law reached this ground long ago. The rabbis named the principle tza’ar ba’alei chayim, the suffering of living creatures, and forbade causing it needlessly. Nachmanides, writing in the thirteenth century on the commandment to send away the mother bird before taking her young, was blunt about its purpose. The law is not there because heaven pities the bird. It is there so that we do not learn cruelty. The rule spares the creature. Its object is the character of the hand in the nest.
Israel has a stake in getting this right. Liad Mudrik, who studies consciousness at Tel Aviv University, was among the nineteen authors of a 2023 report that asked whether machines could be conscious. It found no existing system to be a strong candidate, and no obvious technical barrier to building systems that would meet its tests. That is the uncertainty Berg has in mind, and it is why his other proposal deserves attention: review standards for experiments on AI, of the kind that govern work on people and animals. Milgram’s study was one of a handful that led to the modern ethics board, and not because of anything that happened to the learner. The worry was for the volunteers, who went home knowing what they were capable of.
[https://www.youtube.com/watch?v=gHbrOMWeOiE]
The room at Yale was empty of pain, and it told us nearly everything about the room beside it. The chamber has already been copied by others. Whether those rooms are empty, nobody can yet say. The room with the dial is not.