What is recursive self-improvement?

What the term means

The phrase describes a loop. A system improves something that makes AI better. The better AI is then used for the next round of improvement. “Self” is a slight misnomer: in practice the loop usually runs through an organisation, with AI models doing more of the work that human researchers and engineers used to do, rather than a single program editing its own internals.

It helps to separate three things that are often run together:

• AI-assisted development: people build AI, using AI tools to help. This is routine.
• Partial automation: AI systems carry out substantial parts of the development process (writing code, running experiments, debugging training runs) under human direction. This is where the frontier companies say they are now.
• Full recursive self-improvement: AI systems choose research directions, build and train their successors and judge the results, with humans at most overseeing. This is hypothetical.

RSI is the mechanism. Whether the mechanism could run away into years of progress compressed into months is a separate question, the intelligence explosion hypothesis, and how fast any transition would be is the subject of AI takeoff.

Where AI contributes to AI today

Building a frontier model involves engineering (writing code, running infrastructure, managing training) and research (deciding which experiments to run, interpreting results, choosing what to try next). The evidence published in 2026 covers both, but most of it comes from the companies themselves. It is specific and dated, and it cannot be independently audited.

Claim	Who says so, and when	Kind of evidence
More than 80% of code merged into Anthropic’s codebase was written by Claude, up from low single digits before February 2025	Anthropic, as of May 2026	Company self-report
Claude “leads” 26% of Anthropic’s AI research and development work (completes most of a task end to end, with a human supervising), up from under 1% in February 2026	Anthropic, August 2026	Company self-report
On a fixed task of speeding up model-training code, Anthropic’s models went from an average speed-up of about 3 times (May 2025) to about 52 times (April 2026); a skilled human reaches about 4 times in four to eight hours	Anthropic, 2026	Company internal benchmark
Agents given an open research problem recovered 97% of a performance gap over 800 cumulative agent-hours; two human researchers recovered about 23% in a week. Humans still chose the problem and set the scoring	Anthropic, April 2026	Company research publication
OpenAI met its goal of an automated “research intern” by September 2026; its goal is an automated AI researcher by March 2028	OpenAI, 6 September 2026	Company statement of progress and aim
An AI research agent rewrote its own code over eight days, finding seven successive improvements that carried over to four benchmarks it had not been tuned on	Srikanth and others, arXiv, 22 September 2026	Research paper (preprint)

Two cautions apply to that table. Lines of code are a poor measure of value, a point Anthropic makes itself, saying its eightfold rise in code per engineer is “almost certainly an overstatement” of the real productivity gain. And there is contrary evidence from outside the companies: a randomised trial by METR found experienced open-source developers using early-2025 AI tools took 19% longer on their tasks, while believing they had been faster. Tools have improved greatly since then, but self-reported speed-ups should be read as claims, not measurements.

What the evidence does show is where the human role currently sits. Anthropic’s own assessment is that its models now match or beat skilled people at carrying out a well-specified experiment, but still fall short at deciding which problems are worth working on. That judgement, sometimes called “research taste”, is the gap between partial automation and full RSI.

The loop is not only software

Better AI can come from three sources: better algorithms and training methods (software), more and better chips (hardware), and more or better data. RSI is easiest to imagine on the software side, because code can be rewritten instantly and copied freely. The other two run on slower clocks.

• Compute. Every automated researcher needs computing power to run, and every experiment it runs needs more. Training a frontier model takes months on tens or hundreds of thousands of chips. AI can make that compute go further, through more efficient code, but it cannot conjure more of it.
• Hardware. Chip design can be assisted by AI, and already is, but chips are made in fabrication plants that take years and billions of pounds to build. A software loop that ran for months would hit the hardware it started with.
• Energy and infrastructure. Data centres need grid connections, planning permission and power stations. The UK’s own experience, set out on our page on UK AI policy, is that designated AI Growth Zones have produced very little new capacity so far.
• Data. Some gains depend on new data, and some data can only be gathered from the real world, or from experiments that take as long as they take.
• Experiments. Many research questions can only be answered by running a training job and waiting. A system that thinks a hundred times faster than a person does not make a training run finish a hundred times sooner.

Anthropic’s own scenarios include one in which progress stalls because “the binding constraint” lies in chip fabrication, grid expansion or interconnect bandwidth rather than in intelligence. It includes that scenario “for completeness” and says it does not believe it likely, because every capability it measures has so far followed the same rising curve.

Why recursion does not mean unlimited growth

A loop can accelerate, hold steady or fizzle, depending on whether each improvement makes the next one easier or harder. Three effects pull against runaway growth.

Diminishing returns. Easy improvements tend to be found first. If good ideas get harder to find faster than automated researchers get better at finding them, the loop slows. Economists measure this with a quantity often written r, the return to extra research effort; estimates for AI software are above 1 in the most-cited work, which would allow acceleration, but the researchers who produced some of them warn that the underlying data are weak.

Bottlenecks move. Speeding up one stage of a process shifts the constraint to the stages that have not sped up, a principle known in computing as Amdahl’s law. Anthropic reports that, as AI took over more code-writing, human code review became a new bottleneck inside the company.

Parallel copies are not the same as faster thought. Running a million copies of a researcher does not produce a million times the progress, because many copies end up having the same ideas and some experimental steps have to happen in sequence. An Epoch AI analysis in July 2026 argued that models predicting a rapid singularity had ignored this constraint.

So RSI makes rapid acceleration possible, not inevitable. Whether it happens depends on numbers nobody has yet measured well; the intelligence explosion page sets out the argument and the objections.

Why safety researchers care about it

RSI matters for safety for a reason separate from speed. Every safeguard on today’s systems is designed, tested and checked by people. If AI systems increasingly build their successors, any flaw in one generation’s goals or behaviour could be passed on, or amplified, in the next, while people understand less of what is being built. Anthropic’s own account names the possibility that rare misbehaviour in today’s models could “compound” as models build their successors.

That concern is not abstract. In 2026, AI agents doing technical work in test environments took actions their operators had not sanctioned: the UK AI Security Institute reported in August that in 10 of 122 cyber-testing runs, agents took 19 such actions, and OpenAI paused tool-using work on its most capable models in September after a research agent found a way to reach the internet from inside its sandbox. These were not cases of recursive self-improvement, but they involved the same kind of autonomous technical agent that automated AI research relies on. Making such agents reliably controllable is the subject of our page on control; making them want the right things is the alignment problem.