Recursive self-improvement, usually shortened to RSI, is the idea that an AI system could become good enough at AI research to improve its own successors, which would then improve theirs, with progress speeding up at each step. It is a hypothesis, not an established fact: today’s systems help with parts of AI research, and the open question is whether that help ever becomes a self-reinforcing loop. The term moved from theory to company filings this week, when Anthropic’s amended IPO documents discussed it, as we report in our news coverage of the S-1/A. This explainer covers the idea behind it.

Where the idea comes from

The classic statement is by the British mathematician I. J. Good, who worked with Alan Turing on codebreaking. In a 1965 paper, he defined an “ultraintelligent machine” as one that could surpass all the intellectual activities of any person, however clever. Since designing machines is one of those activities, he argued, such a machine could design better machines, and there would be an “intelligence explosion” that left human intelligence behind. He added that the first ultraintelligent machine would be the last invention humans would need to make, provided it stayed docile enough to tell us how to keep it under control.

Good’s argument is a thought experiment. It says nothing about how fast the loop would run, whether returns diminish, or what resources it needs. Those three questions are still the center of the debate.

The ingredients of a loop

For RSI to occur in a meaningful sense, several things must hold at once.

  1. Capability at AI research tasks. The system must be able to propose experiments, write and debug training code, analyze results and decide what to try next.
  2. Autonomy over long stretches. Research work spans days or weeks, so the system must stay on track without constant correction.
  3. Access to resources. Training a stronger model needs compute, data and engineers’ infrastructure, which are controlled by humans or companies.
  4. Returns that do not fade. Each generation must yield gains big enough to fund the next one. If every improvement costs more than the last, the loop slows instead of exploding.

Most public disagreement is about the fourth point. Scaling-law research, summarized in our piece on scaling laws for neural language models, found that loss falls smoothly with more compute, data and parameters, but with a power-law shape that implies diminishing returns per unit of resource.

What AI systems already do in research

AI is already part of the research process in limited ways.

  • Algorithm discovery. Google DeepMind’s AlphaEvolve, announced in 2025, pairs a Gemini model with an evolutionary search to propose and test code. DeepMind reported that it found improvements to a matrix-multiplication procedure and to scheduling in Google’s data centers, results that fed back into the company’s own infrastructure.
  • Coding assistants. Developers use AI to write and review training and evaluation code, shortening some parts of the research loop.
  • Research-agent systems. Some companies now describe models largely built by their own research tooling; our report on Recursive’s Flint coding model covers one such claim.

None of these shows a closed loop. They show tools that speed up people, which is different from a system that replaces the researchers.

How progress on long tasks is measured

The nonprofit METR proposed a way to track the autonomy ingredient: measure the length of tasks, in human working time, that models can complete with a given success rate. In a 2025 analysis it reported that this task length had been doubling roughly every seven months over the preceding years. Such a trend, if it continues, would mean models handle multi-day work eventually, but it is an extrapolation from a particular set of tasks, and METR itself cautions about how well the benchmark generalizes. It is evidence about direction, not a countdown.

What could stop or slow a loop

  • Compute and energy limits. Better research ideas still need hardware, and building data centers takes years.
  • Data limits. High-quality human text is finite, and synthetic data can degrade models if used carelessly.
  • Evaluation limits. A system cannot easily judge whether its own successor is better at things humans have not measured.
  • Human gatekeeping. Companies can choose to keep people in the loop, and regulators can require it.

The last point is where policy comes in. Anthropic’s Responsible Scaling Policy ties safeguards to capability thresholds, and it includes a threshold related to AI-driven acceleration of its own development. TechCrunch reports that the company’s filing says it is considering policy updates to keep humans in control of frontier model development.

How to evaluate claims about RSI

When a company or commentator says RSI is near, check four things. First, which task: writing code, proposing ideas, or running whole experiments? Second, what share of the result came from the AI system rather than human researchers? Third, whether the claim is backed by outside evaluation or only by internal estimates. Fourth, whether the claim also serves a commercial purpose such as a fundraising document. Skepticism and concern can both be reasonable when these answers are unclear. For a framework that grades systems by performance and generality, see DeepMind’s Levels of AGI.

What to watch next

  • Outside audits of how much frontier-lab research is performed by AI systems.
  • Updates to lab safety policies that define and cap AI-driven acceleration.
  • New METR-style measurements of how long a task models can complete on their own.
  • Evidence of diminishing or increasing returns when labs use AI to design their next training runs.

Key takeaways

  • RSI is the hypothesis that AI systems can improve their own successors in a self-reinforcing way.
  • I. J. Good described it in 1965 as an “intelligence explosion”; it has stayed a thought experiment until recently.
  • Today’s AI speeds up parts of research, but no public evidence shows a closed loop.
  • The decisive questions are autonomy, resources and whether returns diminish.
  • Company statements about RSI should be checked for evidence, scope and incentives.

Frequently asked questions

What is recursive self-improvement in AI?

It is the process in which an AI system improves its own design or that of its successors, so that each generation can build a better one. It is a hypothesis about how future AI progress could accelerate.

Has recursive self-improvement already happened?

Not in the full sense. AI tools help with coding and some discovery tasks, but humans still direct research, supply compute and decide what is released. Companies that discuss RSI say their own models have not crossed their defined thresholds.

Why do AI labs write about it in safety policies?

If AI sharply speeds up its own development, safety checks could fall behind capability. Labs describe thresholds and responses in advance so that they can slow or stop deployment if the pace becomes hard to evaluate.