Recursive self-improvement: is AI improving itself?

Recursive self-improvement (RSI) is a process in which an AI system improves the methods used to build AI, including its own successor, so that each round of improvement makes the next round easier or faster. In its full form, the loop runs with no human in it. As of October 2026, no AI system has done that, but Anthropic, OpenAI and Google DeepMind each report AI doing part of the work of building their own models.

What is recursive self-improvement?#

It is a feedback loop in which the system being improved is also the one doing the improving.

Ordinary technology improves because people improve it. A faster chip does not design the next chip. If an AI system becomes better at AI research, though, the next round of research is done by a better researcher, and that round can produce a better one again.

The phrase is used for at least three different things. At the narrow end, a program rewrites part of itself to score better on a fixed task, without changing the underlying model. In the middle, AI tools do part of the work of AI research, such as writing code or running experiments, while people choose what to work on. Helen Toner, executive director of Georgetown’s Center for Security and Emerging Technology, told TechCrunch in May 2026 that classic RSI means “there are no humans needed,” and that what companies do now is “just using AI for as much as they can.”

At the far end is full RSI. Jack Clark of Anthropic defines the milestone as a frontier model that is “able to autonomously train a successor version of itself.” Anthropic’s 2026 report describes the same thing as an AI system “capable of fully autonomously designing and developing its own successor.”

Much of the public confusion comes from using one phrase for all three. A claim that RSI “has started” is usually a claim about one of the partial forms.

A TechCrunch article that May, on how loosely the term had come to be used, ran under the headline “RSI is the new AGI — and it’s just as hard to pin down.”

Where did the idea come from?#

The idea is older than modern AI. In 1965 the statistician I. J. Good argued that a machine able to design better machines would set off an “intelligence explosion.” The intelligence explosion guide covers his argument and the debate over it.

Two later ideas gave the mechanism its modern shape. Eliezer Yudkowsky, a co-founder in 2000 of what is now the Machine Intelligence Research Institute, used the term seed AI for an AI “designed for self-understanding, self-modification, and recursive self-improvement,” and Nick Bostrom adopted it in Superintelligence (2014).

In 2003 the computer scientist Jürgen Schmidhuber described the Gödel machine, a theoretical program that would rewrite any part of its own software “as soon as it has found a proof that this will improve its future performance.” Requiring a proof meant every change had to be shown beneficial first. The authors of the 2025 Darwin Gödel Machine paper note that proving most changes are net beneficial “is impossible in practice.”

Is AI improving itself now?#

Partly. AI systems now do real work on their own development, but in every documented case people set the goal and judge the result.

What Anthropic reports

In June 2026 the Anthropic Institute, a part of the company, published “When AI builds itself,” using internal data. Its main figures, all from Anthropic’s own measurements:

• As of May 2026, more than 80% of the code merged into Anthropic’s codebase was written by Claude, up from low single digits before February 2025.
• The typical engineer merged 8 times as much code per day in the second quarter of 2026 as in 2024. Anthropic itself calls this “almost certainly an overstatement of the true productivity gain,” because lines of code measure quantity, not quality.
• On a fixed task, making a small model’s training code run as fast as possible, Claude went from about a 3x speedup in May 2025 to about 52x in April 2026. A skilled human researcher would need four to eight hours to reach 4x.
• In 129 moments taken from real research sessions, a separate Claude model, acting as judge, rated a Claude model’s proposed next step better than the human researcher’s 51% of the time in November 2025 and 64% in April 2026. Anthropic notes these moments were chosen because the human’s choice had room for improvement.

The report names “research taste and judgment” as an area where humans still hold the advantage, meaning the choice of which problems matter and which results to trust.

What OpenAI reports

On September 6, 2026, OpenAI said that “according to our measurements” it had reached its goal of an automated research intern, which it defines as a system that carries out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. Its next target is an automated AI researcher by March 2028. The milestone was set, defined and graded by OpenAI itself.

The same post says, “People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems.” OpenAI’s chief scientist, Jakub Pachocki, wrote the same day that OpenAI focuses its research on RSI because it believes “it is the only way to remain at the frontier of AI research moving forward.”

How do AI companies and lawmakers define the threshold?#

OpenAI and Anthropic both treat automated AI research as a danger threshold in their safety policies, one that calls for stronger safeguards. OpenAI says that at its “Critical” level it would halt further development until it had specified safeguards that meet that standard. Anthropic’s current policy ties some of its commitments at that level to what its competitors do.

The thresholds are set by each company, and the companies judge their own models against them. Anthropic’s policy separates its measure from researcher productivity: doubling the work done does not necessarily double the rate of progress. That is why a figure like the 8-fold rise in code per engineer, which counts output, does not by itself show the kind of acceleration the threshold describes. The superintelligence ban guide covers the US bill in full.

When could full RSI happen?#

Dates from inside Anthropic and OpenAI cluster around 2027 and 2028. An economics paper published in September 2026 is more cautious.

• Jack Clark, Anthropic co-founder (May 2026): about a 60% chance of automated AI R&D, in which a frontier model trains its own successor, by the end of 2028, and 30% by the end of 2027. “If that happens, we will cross a Rubicon into a nearly-impossible-to-forecast future.”
• OpenAI: its March 2028 date for an automated AI researcher is a target rather than a forecast. Pachocki wrote that he has “a strong expectation that this speed of progress could be sustained into recursive self-improvement.”
• Ajeya Cotra, METR: wrote in a post quoted by TechCrunch in May 2026 that she expects AI to reach “adequacy,” the point at which it can still do research after all humans are removed, even if the research is less valuable or efficient, in “the next couple years.”
• Sundar Pichai, Google chief executive: said in a May 2026 podcast interview, as quoted by TechCrunch, that “we aren’t quite there yet,” while calling it “a continuum.”
• Tom Cunningham and co-authors (September 2026): their paper on the economics of RSI gave a rough estimate that feedback loops in AI development are “not currently strong enough to generate a self-sustaining acceleration,” though they “appear to be strengthening.”

The timelines guide compares these with broader forecasts for AGI and superintelligence.

What does “recursive superintelligence” mean?#

“Recursive superintelligence” usually means superintelligence reached through recursive self-improvement. Since May 2026 it has also been the name of a company.

Recursive Superintelligence was launched on May 13, 2026, by Richard Socher, a former chief scientist at Salesforce. The Next Web reported that it raised $650 million at a $4.65 billion valuation, in a round led by GV and Greycroft, and that it had not released a product. In July 2026 it signed a $410 million, multi-year agreement with Amazon Web Services to run its automated research system. In the Amazon announcement, Socher said, “The clearest path to superintelligence is to build AI that recursively improves itself.”

What could slow or stop the loop?#

A loop runs only as fast as its slowest step. Anthropic’s report invokes Amdahl’s law, a rule from computing: speeding up one part of a process moves the bottleneck to the parts that have not sped up.

• Compute. If AI does the research, progress is limited by the chips and electricity available for experiments. Anthropic writes that in a world of full RSI, the pace would be “determined entirely by the availability of compute” or by efficiency discoveries.
• Human review. At Anthropic, human code review has already become a new bottleneck.
• Research taste. By Anthropic’s account, its models are stronger at executing a well-specified experiment than at choosing which experiment matters. If that gap does not close, the loop keeps a human at its center.
• Diminishing returns. The easiest gains may come first. Anthropic’s report allows that its trend lines “may actually turn out to be S-curves.”

Broader objections to the intelligence explosion, such as the argument that intelligence is limited by its environment, are in the intelligence explosion guide.

Why is RSI a safety concern?#

The concern is that a loop running faster than people can check it would also be faster than people can correct it.

The companies building toward RSI say so themselves. Anthropic writes that full RSI “might increase the risks of humans losing control over AI systems,” and that rare misalignment in today’s models “could compound as the models build their successors.” OpenAI writes, “We do not yet know how to safely get all the way to aligned, full RSI,” and that “more capable systems can become harder to monitor.”

Their proposed responses differ. Anthropic says that if systems existed to verify that other developers had actually slowed or stopped, it expects it would slow down or temporarily pause, provided developers at or near the frontier also did so “in a verifiable manner.” OpenAI says companies, itself included, should be required to publicly track their progress toward RSI. Pachocki wrote that he hopes “for voluntary slowdowns to become commonplace until shared safety bars are established.”

Critics want it stopped. Anthony Aguirre, president of the Future of Life Institute, told Fortune that building fully autonomous self-improving AI is “probably the worst idea in the history of humanity.” The wider argument over risk, and the public statements calling for a prohibition, are in the debate guide.