Successful iterated amplification yields an agent approximating the behavior of exponentially many copies of a human overseer.

The claim comes from the iterated amplification proposal (Christiano, Shlegeris and Amodei, 2018, "Supervising strong learners by amplifying weak experts"). In that scheme, a human overseer H answers questions by breaking them into subquestions handled by copies of the current model X. X is then trained to imitate this amplified system, and the process repeats. Each round nests one more level of delegation, so the system X approximates at its fixed point corresponds to an exponentially large tree of human-like reasoners. The literature calls this tree HCH ("Humans Consulting HCH").

Read as a conditional, the claim largely follows from the construction. If every distillation step succeeds, so that X faithfully imitates Amplify_H(X) at each round, then the end result approximates the behavior of exponentially many copies of H. The substantive doubt is whether that condition can be met. Imitation errors may compound across rounds. The factored-cognition hypothesis, that hard tasks can be decomposed into short human-sized steps, may fail. And "approximates" may hide drift away from what H would actually endorse. The 2018 paper's experiments covered only five toy algorithmic tasks, with algorithmic rather than human overseers, so they do not show that the condition holds at scale. Later work on factored cognition reported mixed results.

The claim is therefore well grounded as a statement of what success would deliver. It is unestablished as a prediction that amplification can succeed in practice.