Recursive self-improvement, often shortened to RSI, describes a feedback loop in which an AI system helps improve the research, engineering, evaluation, or training of the next generation of AI. In the strongest version of the idea, each improved system helps create a still more capable successor. That could make progress faster than a development cycle driven only by human researchers and engineers.

The term does not mean that a chatbot updates itself after every conversation, or that today’s AI has independently taken over its own development. Anthropic’s September 2026 discussion of the subject says it is delegating a growing share of AI development to AI systems, but also says full autonomous recursive self-improvement has not arrived and is not inevitable. Its examples describe AI assisting with coding, well-specified experiments, and longer-running engineering tasks while people still set goals, review work, and make higher-level judgments.

The loop can be gradual. A model might help write training code, propose experiments, analyze evaluation results, improve tools used by researchers, or generate data for the next training run. Those gains can make the next model-development cycle cheaper or faster, which in turn creates more opportunity for AI assistance. The hard part is not merely building a faster loop. It is knowing whether the system’s improvements are robust, safe, and aligned with the intended goal.

That is why recursive self-improvement is often discussed alongside alignment, model evaluations, and circuit-based interpretability. In his recent essay, Dario Amodei argued that the prospect of AI increasingly helping build future AI systems is a reason to pace capability advances so that safety work and independent evaluation can keep up. That is a policy position, not a settled forecast. What is clear is that AI-assisted development is already changing how frontier labs work; how far the feedback loop can safely go remains uncertain.

If this raises questions about who sets the boundaries when AI systems take on more consequential work, the AI Governance course is a useful next step for learning how organizations can set oversight, accountability, and responsible-use practices around AI.*