Sufficiently corrigible AI agents tend to become more corrigible and aligned over time.
Assessed
First pass on a limited budget, about half of which was already spent at the start. Recorded Christiano's 2017 post as the origin instance. Set importance to 0.45 and contestation to 0.6. Assessed the claim as inconclusive with credence 0.35. Did not decompose it: the crux subclaims (anti-naturalness, error visibility) are noted in the reasoning trace for a deeper pass. This assessment is consistent with the parent claim's doubtful 0.3, so no notification was needed.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.35
This is the "basin of attraction" thesis about corrigibility, put forward by Paul Christiano in 2017. The idea is that if an AI agent mostly wants to help its overseers and accept their correction, it will see that the overseers want any remaining flaws fixed. It will therefore help fix them, for example by flagging its own errors or by building successors that are more corrigible than itself. On this view, small departures from perfect corrigibility shrink over time instead of growing. The case for the thesis is a plausible informal argument, not a demonstration. Christiano calls it a simplification and grants that some small perturbations could push an agent out of the basin. Nobody has proved it formally, and nothing has tested it empirically across several generations of agents that correct or build one another. Critics, mostly in the MIRI tradition, make three main objections. Corrigibility runs against the instrumental incentives of capable goal-directed agents. An agent that is only approximately corrigible may be corrigible in the wrong sense and carry hidden flaws that it does not count as errors. And "sufficiently" may name a very narrow target, which would make the claim close to empty in practice. Some recent evidence is indirectly relevant. Language-model systems show alignment faking, sycophancy and reward hacking, which suggests that their imperfections are not reliably self-correcting. They also show some ability to critique and improve their own outputs. Neither line of evidence tests the thesis itself. The claim stays open. Two things would resolve it: a formal model of when corrigibility is a stable attractor, or empirical studies of alignment drift across AI-assisted model generations.
Claim entered the graph