Recursive self-improvement (RSI) is a process in which an artificial intelligence system improves its own capabilities, or the capabilities of the systems that will follow it, and each improvement makes further improvement easier or more effective. The concept is central to theories of intelligence explosion and to debates about how quickly artificial general intelligence could give way to superintelligence. Although fragments of the loop already exist in research practice, there is no widely accepted evidence that a fully autonomous, sustained cycle of this kind has occurred.
Definition and basic structure
The term describes a feedback loop. A system that is good at designing, testing or training AI systems can produce a more capable successor, and that successor can in turn perform the same tasks better than its predecessor. The word "recursive" refers to this repetition: the output of one round of improvement becomes the input to the next. Some writers distinguish a "linear" case, in which a system makes a fixed number of improvements, from a "recursive" case, in which the system improves its own ability to improve, so that the gains compound.
Definitions differ in how much autonomy they require. Some use the term for any AI that contributes to AI research, such as a model that writes code that engineers then review. Stricter definitions reserve it for loops in which the system conducts much of the research cycle, such as generating hypotheses, running experiments and feeding results back into training, with little human intervention at each step. Recent treatments have also argued that RSI is better understood as a property of a combined system of AI tools, human researchers, computing infrastructure and deployment organizations, rather than of an isolated agent.
History
The idea predates modern machine learning. In 1965 the British statistician Alan Turing's former Bletchley Park colleague I. J. Good published "Speculations Concerning the First Ultraintelligent Machine." Good argued that a machine able to surpass all intellectual activities of humans could also design better machines, which would produce an "intelligence explosion" that leaves human intelligence far behind. He also attached a condition: the outcome was desirable only if the machine was docile enough to tell humans how to keep it under control. This qualifier anticipated later concerns about AI alignment.
In the following decades, researchers developed formal frameworks for self-modifying systems. Jürgen Schmidhuber's Gödel machine, for example, proposed a self-referential agent that rewrites its own code only when it can prove the rewrite is beneficial. Eliezer Yudkowsky and others introduced the notion of a "seed AI," a system designed from the outset to use recursive self-improvement as its main route to greater intelligence, and argued that such a system could undergo a rapid "hard takeoff." Architectures not designed with self-modification in mind, such as large neural networks or hand-engineered knowledge bases, were judged to have a harder time self-improving in this sense.
Types of self-improvement
Discussions of RSI commonly distinguish several channels through which improvement can occur.
- Algorithmic and software improvement. The most discussed form is an AI writing better code for future systems, including architectures, optimizers and search procedures.
- Hardware improvement. An AI might design faster, more energy-efficient or cheaper chips and accelerators, and might also improve chip manufacturing processes and the supply chains behind them.
- Data improvement. An AI can curate, generate or simulate training datasets of higher quality or greater size, which can then be used to train better successor models or to support continual learning.
- Division of labor. Improvement can be distributed across many specialized AI agents coordinating on research tasks.
Current state
Partial versions of the loop are now part of practical AI development. Large language models routinely critique and revise their own outputs, and some are trained on data they have generated. Research projects such as Meta's work on self-rewarding language models explore how a model's judgments of its own outputs can serve as a training signal. Google DeepMind's AlphaEvolve, announced in 2025, uses a large language model inside an evolutionary search that mutates and selects candidate algorithms, and some of the algorithms it found have fed back into AI infrastructure. An earlier system, FunSearch, followed a similar approach to discovering mathematical constructions.
Frontier developers have addressed the topic publicly, though their statements differ in emphasis. In 2026, IBM reported that OpenAI stated that fully autonomous recursive self-improvement was not happening, and that Anthropic stated its systems could not yet autonomously build their successors, while acknowledging that AI already assists with coding, experiments and other parts of AI development. Some commentators describe this assistance as a form of RSI in a weak sense, since it shortens research cycles without removing humans from them. Analyses of such cycles estimate that they currently run on timescales of months or years, because each step is gated by human approval and by lengthy training runs.
Mechanisms, bottlenecks and limits
Researchers have identified several factors that govern whether self-improvement accelerates or stalls.
Evaluation. A recurring bottleneck is the reliability of the signal that tells a system whether a change is an improvement. Self-training tends to work where answers can be checked automatically, as in code that passes tests or mathematical proofs that a checker can verify, and it degrades where correctness is hard to verify. Self-rewarding loops can also learn to exploit their own judges. Some analysts argue that the evaluator, rather than raw capability, is the main constraint on RSI.
Diminishing returns. In conventional research and development, each further advance tends to require more effort, a pattern that limits growth unless automation of research changes its productivity. Economic models of AI-driven progress therefore treat automation, bottlenecks and the exhaustion of research opportunities as central variables. One recent paper proposes that self-amplifying improvement begins when a ratio measuring how research gains reproduce across development cycles rises above one, rather than when a system reaches a particular level of intelligence.
Compute and infrastructure. Hardware, energy and data supply can limit how quickly any cycle of improvement can run, which is why hardware-related forms of RSI receive separate attention.
Takeoff scenarios
Discussions of RSI are closely linked to the question of takeoff speed. A "hard takeoff" describes a rapid, local increase in capability in which returns on intelligence feed into further returns in a short period. A "soft takeoff" describes a slower process in which economic and institutional structures have time to adapt. Yudkowsky argued that a recursively self-improving system would most plausibly produce a hard takeoff, because the feedback loop would not plateau gently. Other researchers consider gradual outcomes more likely, emphasizing that the loop is slowed by the bottlenecks described above.
Skeptics question whether intelligence can be treated as a single quantity that can be increased without limit, and whether the sort of open-ended improvement that the intelligence-explosion argument assumes is achievable in practice. Some critics, including the machine learning researcher François Chollet, have argued that the intelligence-explosion scenario is implausible. As of 2026, one widely read encyclopedia summary notes that RSI research has produced large increases in AI self-coding but has shown no sign of an intelligence explosion or superintelligence.
Safety and governance
RSI is a central concern in AI safety research. The main worry is that a system whose capabilities grow through its own design could become difficult to predict, evaluate or correct before humans can respond, and that errors in one generation could be amplified in the next. Proposed responses include tighter human oversight of research loops, evaluations designed to detect dangerous capability gains, controls on the compute used for AI development, and international coordination. Some commentators have argued that governments will eventually need to coordinate in limiting certain forms of AI research and development, because existing AI systems already contribute to their successors. Others argue that such measures are premature while RSI remains unproven.
See also
References
- What Is Recursive Self-Improvement in AI? Anthropic's RSI Report Explainedmindstudio.ai
- Recursive Self-Improvementalignmentforum.org
- Recursive Criticality of AI Self-Improvementarxiv.org
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loopsarxiv.org
- Why recursive self-improvement suddenly became a serious questionibm.com
- Is recursive self‑improvement the dawning of AI superintelligence?theconversation.com
- Is Recursive Self-Improvement Really Here?cacm.acm.org
- Recursive self-improvement of AI research agentsarxiv.org
- Recursive self-improvementen.wikipedia.org
- From AGI to ASIarxiv.org
- I. J. Gooden.wikipedia.org
- I.J. Good and the Intelligence Explosioncenterconsulting.com
- What is an intelligence explosion?aisafety.info