AI Self-Improvement Levels Framework Identifies 5 Stages

Over 30 researchers published a 75-page paper defining a 5-level framework for AI self-improvement, from Level 1 (human decides what to improve) to Level 5 (AI rewrites the process of creating next-gen AI). A review of 491 papers found 75% at Levels 1-2 and less than 6% at Level 5.
Key insights
Five Levels of AI Self-Improvement Defined
The framework categorizes AI self-improvement into five levels: Level 1 (human decides what to improve), Level 2 (AI assists in improvement), Level 3 (AI proposes improvements), Level 4 (AI autonomously improves within constraints), and Level 5 (AI rewrites the process of creating next-generation AI). The paper reviewed 491 research papers and found that 75% are at Levels 1-2, while less than 6% reach Level 5.
Most Current Research Stays at Low Self-Improvement Levels
The review indicates that the vast majority of AI self-improvement research is still at early stages where humans play a dominant role. Only a small fraction of papers demonstrate AI systems that can autonomously redesign the improvement process itself, highlighting the significant gap between current capabilities and full recursive self-improvement.