Work

From Recursive Self-Improvement to Human-AI Reciprocal Improvement

Year: 2026
Type: preprint
Abstract: Recursive self-improvement advances most readily in domains where results can be verified reliably, such as mathematics and coding. Beyond these domains, progress slows as reliable signals for evaluating and guiding improvement become harder to obtain, resulting in a jagged frontier of AI capabilities. Yet even where AI self-improves autonomously, its growing capabilities do not necessarily translate into greater human ability to understand, use and build upon its outputs, challenging both continued human participation and effective oversight. We argue that recursive progress should therefore also encompass the growth of human capabilities. We propose Recursive Reciprocal Improvement (RRI) as a process of human–AI co-evolution, in which AI advances help people develop their abilities to propose worthwhile problems, judge results and contribute to higher-level problem solving, while these evolving capabilities inform subsequent generations of AI. Human capabilities thus become both source and target of recursive progress, with AI and human development mutually reinforcing.
Language: English
Open Access status: green
OpenAlex ID: W7236807578
Cites:
Cited by:
Related to: 10