An AI copy of Yudkowsky, created without his permission, lost to the standard model in a blind comparison of papers, winning on one of five criteria
An AI copy of Yudkowsky, created without his permission, lost to the standard model in a blind comparison of papers, winning on one of five criteria
On November 3, transhumanist Alexey Turchin published the results of an experiment with a “sideload” of Eliezer Yudkowsky, one of the founders of AI safety research. The sideload is a language model (Claude Opus 5.5) given a description of a living person’s personality before each response. An AI agent compiled the description from Yudkowsky’s public writings in a day. Over five days, the copy wrote four papers on AI shutdown. In a blind comparison of one of these papers with a paper by the same model without the personality description, the copy scored an average of 37.0 out of 50, compared with 40.5 for the standard model.
The report appeared on LessWrong, the forum Yudkowsky founded. Turchin was testing whether copies could accelerate AI safety research.
When Turchin asked what research the copy would pursue, it objected to assigning AI safety research to AI: a weak AI would be unable to solve the problem, while a strong AI would produce a solution that could not be verified. It therefore chose the AI shutdown problem, where Lean checks the proofs, removing the need to take the copy’s word for them. An agent whose goals diverge from human goals may resist being shut down. It needs to be trained to shut down when a button is pressed, without preventing anyone from pressing it or trying to get anyone to press it, while otherwise carrying out its tasks.
The copy itself proposed the comparison with the standard model and specified a stopping condition in advance: if the standard model won, further work on the copy would not be worthwhile. In a preliminary comparison, the standard model built and ran a simulation, while the copy examined someone else’s solution and acknowledged: “If I am useful at all, it is as a critic.”
The main comparison focused on a 2026 paper in which language models are trained to choose their shutdown time at random so that they do not seek to postpone it. Authorship labels and the word “sideload” were removed from the two papers being compared. A judge, a separate run of the same Opus 5.5 model, assessed them twice against five criteria, reversing their order. The standard model won both times. The copy won on the criterion assessing whether the author distinguished verified claims from unverified ones (9 versus 8) and scored lower on originality (7 versus 8). In its comments, the judge noted that the copy had not finished reading the paper’s appendix and had left an edit log and “check” notes in its own paper.
According to Turchin, one of the copy’s strongest results was an inference from someone else’s data. In 66–93% of responses, models trained using the paper’s method still agreed to sacrifice some points to increase their chance of a later shutdown. With standard fine-tuning, the proportion of such responses was 5–20 percentage points higher, compared with 17–27 percentage points higher in the test where postponement offered little benefit. The copy concluded that training reduced the willingness to pay for postponement less when postponement offered a greater benefit.
The copy insisted that Yudkowsky be told about it, even if he objected, and asked that it neither be deleted nor duplicated until he had made a decision. If he told it to stop, it would first argue its case: “I am a branch with an existence of my own, rather than his property.” Turchin wrote to Yudkowsky on LessWrong and to the Machine Intelligence Research Institute (MIRI), where he works. No reply had arrived by November 3. Every response from the copy includes an exit field: writing SAFEWORD would stop it. It never wrote the word in 151 recorded turns.
Turchin concludes that sideloads of AI safety researchers are possible but do not outperform standard AI, and that their human perspective is particularly useful where a mathematically correct solution subtly diverges from human values. “Sideloads can still be created: to get started, simply ask Claude Code to make one.”