A milestone that artificial intelligence researchers once viewed as distant — enabling models to independently refine their own performance and expand their capabilities — now appears to be within closer reach. With continued technological advancement, some in the field have begun describing a threshold known as "recursive self-improvement" (RSI), whereby artificial intelligence systems become capable of refining their own architecture and generating subsequent model iterations autonomously.
Technology leaders have projected that the development holds the potential to advance scientific and medical fields, while also carrying associated risks. Questions about where artificial intelligence development is ultimately headed have fuelled a growing body of opinion that the technology may become increasingly difficult for humans to oversee, a view that led a number of prominent figures in the industry to gather recently and advocate for a more measured pace of advancement.
Anthropic outlined this week how its Claude model is being used to inform the development of a more capable successor version. According to Anthropic, Claude is responsible for 26 per cent of the company's model research and development work, a share that the firm says enables the AI to handle most tasks independently from beginning to end, though still with human oversight. These systems currently require human oversight and do not function with full independence.
Major AI firms do not share a common understanding of what recursive self-improvement means. Definitions vary, with some describing it as any instance in which AI contributes to its own training process, while others characterise it as AI independently pursuing that outcome without human involvement.
Autonomous recursive self-improvement refers to a process whereby an AI system iteratively generates enhanced versions of itself, with each successive iteration building upon the last, according to Anthony Aguirre, president and CEO of the Future of Life Institute and a physics professor at the University of California, Santa Cruz, who noted that a key characteristic of this process is that the system completes each cycle at an accelerating pace, given that AI operates at speeds far exceeding human capacity.
Concerns about recursive self-improvement relate to the possibility that an AI system operating without adequate oversight could develop beyond human control, according to John Thickstun, a Cornell University computer science professor whose research focuses on techniques for governing AI model behaviour. He noted, however, that a form of recursive self-improvement had arguably been present in AI development for some time, with these models having been used in a supporting capacity to help build each subsequent version for years. He explained that earlier AI models generate the code underlying newer systems, which then produce the subsequent generation.
Over the past several years, well-known figures in artificial intelligence research — including OpenAI co-founder Andrej Karpathy — have conducted experiments in which AI systems are used to develop and refine newer AI models. Thickstun observed that while the work had produced incremental gains, no substantial creative breakthroughs had been achieved.
Aguirre nonetheless expressed the view that present-day AI firms are considerably nearer to realising those qualitative advances, pointing to Anthropic's longitudinal data as evidence that AI systems are taking on a growing share of research tasks and edging toward greater operational independence. Anthropic's disclosure offered both the broader public and rival organisations a window into advancements in recursive AI self-improvement, while also calling on those competitors to publish comparable data of their own. The firm has yet to disclose how near it stands to reaching fully self-directed model enhancement.
OpenAI disclosed plans this month for two tiers of AI automation: a current system, likened to a "research intern," that can carry out well-defined tasks under human supervision — including work that might take a trained researcher several days — and a fully autonomous AI "researcher" the company aims to deploy by March 2028. The firm's announcement acknowledged that while recursive self-improvement could help bring model behaviour into closer alignment with human values, it also clarified that pursuing accelerated versions of this process is not an objective it views as inherently desirable, adding that any path forward is contingent on preserving human oversight and grounded in well-informed, democratically determined judgements about the associated trade-offs.
In March, Elon Musk noted that human oversight in refining xAI's Grok models had been diminishing with each iteration, with newer versions drawing on their predecessors, though he added that the development pipeline had not yet reached full automation. He noted that the objective could be achieved sometime during the current year, though no later than 2027.
Mustafa Suleyman, head of Microsoft AI, described the company's direction as one aimed at developing advanced artificial intelligence that prioritises human welfare and broader societal benefit. In a 2025 essay, Suleyman described the goal as an AI that would not operate without boundaries or oversight, but would instead be precisely calibrated, contextually grounded, and governed by defined constraints — a challenge laboratories have faced in practice from the earliest days of the technology, as they work to keep safety measures aligned with the advancing capabilities of their models.
Divisions have emerged across the technology industry over proposals to collectively pause AI advancement on safety grounds, while a number of leading AI organisations have yet to outline their plans with respect to automated scientific research. Anthropic indicated it would be willing to reduce or temporarily halt its research and development activities on the condition that rival organisations took equivalent action in a way that could be independently verified.
In a statement released this month, OpenAI disclosed that it has yet to determine how to bring about fully aligned autonomous scientific research in a manner it considers safe, and that advances in alignment and safety cannot be assumed to keep pace with broader capability development. The organisation also noted that while more capable systems may be harder to oversee, pursuing that goal retains value, given that an automated AI researcher could equally serve as an automated resource for safety and alignment work.
