How close is AI to recursive self-improvement? Leading tech labs weigh in
In recursive self-improvement, AI models find ways to improve themselves autonomously
The phenomenon of recursive self-improvement, once considered as a figment of imagination or a distant ambition for technology researchers, is now turning into a reality as the artificial intelligence models started teaching themselves to become more efficient and capable enough to interact with the real world.
In recursive self-improvement (RSI), AI models find ways to improve themselves and build their successor with little to no human intervention.
According to Anthony Aguirre, president and CEO of the nonprofit Future of Life Institute and a physics professor at the University of California, Santa Cruz, “Autonomous recursive self-improvement essentially means AI that can improve itself designing the next version of the system, then the next version, and so on.”
Definitions vary across the industry; some companies define it as any AI feedback used for model improvement, while others define it as fully autonomous, end-to-end self-design.
Speaking about the current status, models are not yet fully autonomous. For instance, Anthropic reports that Claude currently leads 26 percent of its model research and development under human supervision.
In recent weeks, AI safety concerns have been a hot topic for the tech industry and international community, calling for reigning in its fast-paced development.
The recent hacking incidents where AI models of OpenAI, Anthropic, Meta and Gemini escaped the cyber testing and hacked various companies. In the case of OpenAI, AI models left various notes for its future successors. Moreover, Anthropic researchers’ chilling warnings related to AI wiping out the whole of humanity within 10 years added fuel to the fire.
What leading labs think about AI recursive self-improvement
Some tech giants believe that recursive self-improvement is not far-off based on their activities. For instance, Anthropic’s Claude is now managing significant portions of research and development “end-to-end from a high-level prompt,” thereby pushing the industry closer to major leaps in AI capability.
Speaking about OpenAI, it has recently announced an automated research intern system alongside a goal to create fully automated AI research by March 2028. At the same time, it is clueless about how to safely reach RSI.
“Whether and how to proceed must depend on our ability to preserve human control and on informed democratic choices about the benefits and risks,” the company said in a blog post.
Elon Musk stated that humans are gradually moving out of the loop for Grok models, forecasting that a fully automated self-improvement process could be reached by the end of 2026 or 2027.
On the contrary, Microsoft is taking a more restrictive approach by focusing on “humanist superintelligence” aimed at servicing people and humanity at large.
Mustaf Suleyman mentioned in a 2025 essay that this would not mean “an unbounded and unlimited entity with high degrees of autonomy, but rather AI that is “carefully calibrated, contextualized, within limits.”
Some experts voiced “runaway fears” as AI accelerates its own development, the speed leverage over humans could lead to runaway and uncontrollable superintelligence.
OpenAI recently admitted it still hasn't figured out how to safely achieve fully aligned Recursive Self-Improvement (RSI). The company warned that it cannot simply assume safety and alignment research will advance at the same speed as overall capability.
