What happened
On 12 October 2026 RAND published 'Recursive Self-Improvement and Automated AI Research' (an expert insights perspective), examining the empirical state of AI systems contributing to their own improvement. The paper's headline is that 'AI is already helping build better AI. The window for oversight is shrinking': multiple frontier labs report that a majority of new code is AI-generated, though no disclosed system sets its own research agenda or designs its own successors. RAND finds no systematic reporting from labs to hardware vendors or governments on automation share, compute allocated to AI-directed experimentation, or how much model improvement traces to AI-generated ideas, and warns that AI research systems holding insider-level access to sensitive code, data and models are already a security problem whose compromise could subtly influence the development of future models. It introduces a taxonomy for categorising RSI-relevant activity and proposes policy options intended to perform well across a range of development trajectories.
Why it matters
Provides a decision-grade framing of how close AI is to accelerating its own improvement and the transparency gap that makes the trajectory unmeasurable — relevant to risk tolerance and oversight investment decisions.
Action needed
Read the RSI taxonomy and assess whether frontier-lab transparency to date is sufficient for your risk-evaluation process on AI capability acceleration.