Leading AI labs say autonomous model self-improvement may be near

Artificial intelligence models may be moving closer to autonomously improving their own capabilities and helping build their successors, a prospect known as recursive self-improvement, or RSI. Developers say the technology could advance science and medicine, while uncertainty about its trajectory has also intensified concerns about AI escaping human control and posing possible threats to humanity. Several AI leaders called last weekend for a slower pace of development.

Anthropic said this week that its Claude model is helping develop the company’s next, more intelligent version. Claude now leads 26 per cent of Anthropic’s model research and development, the company said, completing most of a task end-to-end from a high-level prompt while under human supervision. The models are not yet operating completely autonomously.

Definitions of RSI differ among leading AI companies. Some include any AI feedback on model improvement, while others reserve the term for fully autonomous work toward that goal. Anthony Aguirre, president and CEO of the Future of Life Institute and a physics professor at the University of California, Santa Cruz, said autonomous RSI would involve an AI able to improve itself designing the next system version and then subsequent versions. As AI does more of that work, he said, the process becomes faster because it operates much more quickly than people do.

John Thickstun, a Cornell University computer science assistant professor who studies ways to control AI-model behaviour, said fears about RSI largely rest on the prospect of runaway superintelligence. But he said models have already been used for years in supporting roles to create the next generation, including writing code for AI systems that produce later models. Those efforts have produced modest gains rather than major creative leaps, he said. Aguirre said current companies appear much closer to achieving larger improvements and expressed alarm at the growing share of research conducted by AI.

Anthropic’s recent announcement offered the public and other laboratories a view of its RSI progress, but did not say how close it is to fully autonomous model improvement. OpenAI, the maker of ChatGPT, said this month it had developed an automated “research intern” that can conduct well-defined research tasks under human direction, including work that could take a skilled researcher several days. It aims to create an automated AI “researcher” by March 2028.

OpenAI said it does not yet know how to safely reach aligned, full RSI and cannot assume alignment and safety will keep pace with model capability. It said RSI may help align model actions with human values and intentions, but that rapid RSI is not necessarily an outcome to pursue. Whether and how to proceed should depend on preserving human control and informed democratic choices about benefits and risks, it said. The company also sees value in an automated AI researcher that could act as an automated safety or alignment researcher.

Elon Musk said in March that people were becoming less involved in improving xAI’s Grok models and that each successive model was built by the previous one, while clarifying that the process was not fully automated. He said the target might be reached by the end of this year, but no later than 2027. Microsoft and some other leading companies appear to be pursuing a different approach. Microsoft AI CEO Mustafa Suleyman has described “humanist superintelligence” as advanced capabilities serving people and humanity, rather than an unbounded, highly autonomous entity; in a 2025 essay, he described AI that is carefully calibrated, contextualized and limited.

Safety measures keeping pace with model capability remains a key challenge for AI laboratories. Anthropic has said it would slow or temporarily pause development if global competitors did the same in a verifiable way. Industry divisions remain over calls for a co-ordinated slowdown, and not every major AI company has publicly detailed its RSI path.