What is recursive self-improvement in AI?

The idea is older than the labs that now discuss it, but 2026 gave it a concrete form: frontier labs using their own models to do AI research, and the question of what happens when the models doing the research are the ones being improved. The Pacing the Frontier statement does not claim this is happening at a dangerous rate now. Its argument is that if it did, the tools to notice and to slow it would need to exist already, and they do not.

The term is often used loosely for any model that writes code, which it is not. A coding agent improving a product is automation. Recursive self-improvement, in the sense the statement uses, is the model improving the process that produces the next model. Whether the distinction can be measured from outside a lab is one of the questions the Slowdown Docket keeps open, and one reason the Amodei essay's commitment to give evaluators standing access matters more than any single number.

Sanders and Casar's Ban Artificial Superintelligence Act reaches the same concern from the other end, by defining superintelligence in terms of capability and proposing to ban it. The three documents disagree about what to do and agree about what they are worried about.