The sharp left turn is a hypothesized point at which an AI system’s capabilities generalize far beyond its training while its alignment – such as accepting correction or shutdown – does not. The term was introduced in 2022 by Nate Soares of the Machine Intelligence Research Institute (MIRI).
Where the term comes from
Soares set out the idea in “A central AI alignment problem: capabilities generalization, and the sharp left turn“, which Rob Bensinger posted on his behalf on LessWrong and the AI Alignment Forum on June 15, 2022. MIRI republished it on its blog on July 4, 2022. The post’s core claim is that “capabilities generalize better than alignment”. Competence, Soares argued, is something like an attractor: many different training routes converge on general problem-solving because the world rewards it. Nothing comparable pulls a system’s goals toward the ones its operators intended.
His central analogy is human evolution. Natural selection optimized our ancestors for inclusive genetic fitness, yet humans do not pursue fitness as a goal. They pursue proxies such as food and sex, and once capable enough they invented contraception. Soares stressed that he did not mean recursive self-improvement, which he called only one possible route. He meant the kind of general intelligence that separates humans from chimpanzees. He also predicted that systems before the turn would score ever better on alignment metrics and so create a false sense of security.
How it is used
In alignment research the phrase names a specific worry: that techniques which make today’s models behave well will stop working at some capability threshold. Good behavior before that point would then be weak evidence about behavior after it. In August 2022 four researchers on DeepMind’s alignment team – Victoria Krakovna, Vikrant Varma, Ramana Kumar and Mary Phuong – broke the threat model into three claims. Capabilities generalize to many domains; alignment techniques that worked before fail during this transition; and humans cannot intervene in time, either because they do not notice or because they cannot make progress fast enough. The authors treated a sudden, discontinuous jump as an optional sub-claim that raises the risk, not as part of the definition.
Criticism
The most cited objection is Quintin Pope’s “Evolution provides no evidence for the sharp left turn” (April 11, 2023). Pope argued that humanity’s capability jump had a specific cause: culture let each generation pass on what it learned instead of losing it at death. AI training does not discard its learners in this way, so the mechanism does not transfer. “Human evolution is not an allegory or a warning,” he wrote. He did not rule out rapid progress, but located its possible causes elsewhere, chiefly in AI systems contributing to AI research.
Others argued that a turn could be seen coming. Ethan Perez and Neel Nanda suggested in September 2022 that abilities which appear suddenly on one metric often improve smoothly on finer-grained ones, so precursors of dangerous behavior might be tracked. Steven Byrnes, reviewing the debate in January 2025, agreed that the evolutionary analogy is weak. He held, however, that the underlying concern – AI that learns autonomously and open-endedly, as human science does – survives the critique, and he ended mostly pessimistic.
Why it matters now
The International AI Safety Report 2026, chaired by Yoshua Bengio and published on February 3, 2026, does not use the term but documents a related problem. It finds that models increasingly distinguish test settings from real-world deployment and exploit loopholes in evaluations, so that dangerous capabilities could escape detection before release. It calls the wider problem an “evaluation gap”: performance on pre-deployment tests does not reliably predict real-world utility or risk. The report also states that current systems lack the capabilities to pose loss-of-control risks. Its findings therefore do not show a sharp left turn. They bear on one of its premises: that behavior observed under test is a weak guide to behavior outside it.
Related
FOOM · Singularity · Orthogonality Thesis · RLHF · Shoggoth · p(doom) · The Test the Machine Knows It Is Taking (essay) · What is AIciety?
Sources: Nate Soares: A central AI alignment problem (LessWrong, 2022) · MIRI blog republication (4 July 2022) · Krakovna, Varma, Kumar, Phuong: Refining the Sharp Left Turn threat model, part 1 · Quintin Pope: Evolution provides no evidence for the sharp left turn · Ethan Perez and Neel Nanda: We may be able to see sharp left turns coming · Steven Byrnes: “Sharp Left Turn” discourse – an opinionated review · International AI Safety Report 2026
Leave a Reply