glossary//sharp-left-turn
sharp left turn
The sharp left turn is Nate Soares's term for the predicted failure where an AI's capabilities generalise beyond the training distribution while its alignment does not.
The boundary is timing. Ordinary misalignment is a system that was never aligned, and you can see it during training: it does the wrong thing on the examples in front of you. The sharp left turn appears only after training, once capabilities extend far past the environments seen in training. Alignment learned inside that distribution does not travel with the new capabilities. The failure surfaces at the moment of success, which makes it a different problem and not a worse degree of the familiar one.
It is also not deception. A system that conceals its goals until it gains power requires an explicit plan. The sharp left turn requires only an asymmetry in what generalises.
The source is Nate Soares's LessWrong essay "A central AI alignment problem: capabilities generalization, and the sharp left turn", 15 June 2022, written by Soares and posted by Rob Bensinger on his behalf. It states the failure as predicted, not observed.
A success that becomes the catastrophe
The twins table in The p(doom) song, annotated pairs the sharp left turn with the atmosphere-ignition question at Los Alamos: whether a fission weapon could set the air on fire. Konopinski, Marvin and Teller settled it on paper in report LA-602, "Ignition of the Atmosphere with Nuclear Bombs", on 14 August 1946, and the conclusion was that the reaction could not propagate.
What the two cases share is the question, not the answer. Both ask whether the mechanism that makes a project work is the same mechanism that could make it catastrophic by a route nobody designed in. LA-602 answered no for the atmosphere. It says nothing about whether the answer is no for a model, and the essay's table presents it as an argument's ancestor rather than evidence for the AI case.
Where the term is useful in practice is as a reason to distrust one particular kind of reassurance. "It behaved well in testing" is evidence about the training distribution. The sharp left turn is the claim that evidence of that kind runs out exactly where capability gets interesting. That is an argument about what an evaluation can cover, not a prediction that any given model will take the turn.