AI-Assisted AI Development

This exploratory report examines whether, how, and under what conditions artificial intelligence systems could contribute autonomously or semi-autonomously to the development of more capable AI systems. The topic matters because even partial acceleration of AI research and development could affect safety evaluation, governance, competition, infrastructure concentration, and societal preparedness. The report does not assume that fully autonomous recursive self-improvement is already possible. Instead, it maps present-day evidence, emerging mechanisms, plausible scenarios, and major uncertainties.

The analysis distinguishes established forms of AI-assisted software development, AutoML, hyperparameter optimization, and evaluation from more tentative evidence on agentic machine learning engineering and research-task automation. It develops four scenarios: present-day AI-assisted development, near-future agentic workflow acceleration, medium-term scaled AI R&D acceleration, and a high-impact but uncertain scenario of self-reinforcing AI development. The report also assesses risks including reliability degradation, AI R&D pipeline security, benchmark gaming, safety-evaluation lag, compute and capital concentration, governance and auditability gaps, open-weight diffusion, dangerous capabilities, and recursive improvement scenarios.

The main finding is that AI-assisted acceleration of parts of AI development is plausible and already partly observable, while strong claims about autonomous end-to-end AI research or rapid recursive self-improvement remain uncertain and scenario-based. Current preparedness institutions are emerging, but their adequacy remains an open question.