The idea that models of artificial intelligence (AI) Learning for themselves autonomously to become more efficient and capable was once a distant ambition for technology researchers, but now it seems increasingly closer to reality.
As technology advances, developers say AI is approaching “recursive self-improvement” (RSI), in which AI models find ways to improve themselves and create their successor. This could bring with it the promise of advances in science and medicine, according to technology company executives, but also risks.
Uncertainty about where all this could lead is at the heart of growing fears of AI escaping human control and potential threats to humanity, prompting several AI industry moguls to join a call last weekend to slow the pace of the technology’s growth.
Anthropic explained this week how its Claude model helps the company develop the next, smarter version of itself. Claude now leads 26% of Anthropic’s model research and development, which the company says means he can complete most of a given task “from start to finish from high-level instruction,” albeit still under human supervision. The models do not work completely autonomously; at least, not yet.
Here are some key points about recursive self-improvement.
1/12 | How to create an image using Microsoft artificial intelligence: Copilot. Among the capabilities of Microsoft’s artificial intelligence, Copilot, is the ability to generate images like this according to the instructions and needs of the users. – artificial intelligence
What is recursive self-improvement?
Major AI companies have different definitions of recursive self-improvement. Some define it as any feedback from the AI about improving the model, while for others it implies that the AI works to achieve that goal completely autonomously.
Autonomous recursive self-improvement is, in essence, an AI capable of improving itself by designing the next version of the system, then the next, and so on, explained Anthony Aguirre, president and CEO of the nonprofit Future of Life Institute and a professor of physics at the University of CaliforniaSanta Cruz.
“The really important thing here is that as AI performs more of these types of tasks, it gets faster, because AI works much, much faster than humans,” he said.
The fear around RSI is largely based on the emergence of uncontrolled superintelligence, said John Thickstun, an associate professor of computer science at Cornell University who researches methods to control the behavior of AI models. However, he added that a more realistic view suggests that a kind of recursive self-improvement has been occurring for some time in AI development.
“We have been using these models for several years in supporting roles to create the next version of them. So, people use the previous generation of models to write code for AI systems that, in turn, create the next generation,” he explained.
For years, prominent artificial intelligence researchers, such as the co-founder of OpenAIAndrej Karpathy, have experimented with the idea of AI models training and improving new AI systems. Those efforts have brought minor improvements, but not major creative breakthroughs, Thickstun said.
Today, however, AI companies are much closer to achieving those breakthroughs, Aguirre said.
“In these Anthropic graphs you can see, over time, that AI is carrying out more and more research and that it is getting closer and closer to being fully autonomous,” he said. “And the result of that success, ultimately, is something that I think is extremely terrifying. I think doing this is probably the worst idea in the history of humanity. And yes, they’re doing it.”
Some AI labs say RSI is not far away
Anthropic’s recent announcement offered the public—and other labs—some perspective on progress in IIA, and encouraged its competitors to share similar metrics. However, the company has not expressly indicated how close it is to achieving a fully autonomous model upgrade.
OpenAI, creator of ChatGPTannounced this month that it has developed an automated “intern researcher,” which it defines as a system capable of carrying out well-defined research tasks under the supervision of a human being, such as “tasks that would take a qualified researcher several days.” The company has stated that it is moving forward with the goal of creating an automated AI “researcher” by March 2028.
The company noted in that statement that while RSI can help align models’ actions with human values and intentions, that does not mean that “rapid RSI is necessarily an outcome we should pursue.”
“The decision of whether and how we should move forward should depend on our ability to preserve human control and democratic decision-making based on information about benefits and risks,” the company wrote in a blog post.
Elon Musk He seems more willing to move on. He noted in March that, in the case of xAI’s Grok models, “humans are less and less involved in the process” of model improvement and that “each successive model is built on the previous one,” but clarified that the process was not yet fully automated. That goal could be reached by the end of this year, he added, “but not after” 2027.
Microsoft and other leading AI companies appear to be taking a different approach.
Mustafa Suleyman, CEO of Microsoft AI, noted that the company is moving towards “humanistic superintelligence,” that is, advanced AI capabilities that are at the service of people and humanity at large. Suleyman wrote in a 2025 essay that this would not mean “an unbounded, unconstrained entity with high degrees of autonomy,” but rather an AI “carefully calibrated, contextualized, and within bounds.”
How discussions about slowing development could affect IHR
A fundamental challenge that laboratories face – and have essentially faced since the beginning of this technology – is ensuring that their security measures evolve at the same pace as the capabilities of the models.
Divisions have emerged in the tech sector over calls for a coordinated slowdown of AI over security concerns, and not all major players in the AI space have spoken specifically about the way forward on RSI.
Anthropicwhich has been one of the most prominent voices in calling for the pace to be moderated, has stated that it would slow down or temporarily suspend its development work, as long as its global competitors did the same, and in a “verifiable” way.
OpenAI explicitly stated this month that it does not yet know how to “safely achieve full, aligned RSI,” adding that the company “cannot take for granted that advances in alignment and security will keep pace.” More capable systems may be more difficult to monitor, he continued, but pursuing RSI is still a goal he says he values because “an automated AI researcher can also be an automated security or alignment researcher.”