TL;DR: AI helps build better AI now, but fully autonomous self-upgrading remains ahead

Recursive self-improvement means an AI system taking a growing role in designing its successor, and leading labs say partial versions are already in use while full autonomy has not arrived. According to The Associated Press, Anthropic reported on Sept. 18, 2026 that its Claude model now leads 26% of its model research and development under human supervision. According to IBM Think, OpenAI says fully autonomous recursive self-improvement is not happening today, and Anthropic says its systems cannot yet autonomously build their successors.

Anthropic's 26% milestone puts the debate in concrete terms

The newest verified fact is Anthropic's disclosure this week that Claude completes much of an assigned research task from start to finish after receiving a broad prompt, while people still oversee the work. According to The Associated Press, the company described that contribution as leading 26% of its model research and development.

That figure gave outside researchers and rival labs a rare public measure of how much model development is machine-assisted. According to The Associated Press, Anthropic asked competitors to publish comparable numbers, but it did not state a date for reaching fully autonomous improvement.

The disclosure landed days after executives from several frontier companies joined a public call last weekend to slow the pace of AI growth. According to The Associated Press, concern that AI could evade human control helped motivate that statement, with recursive self-improvement central to the worry.

For U.S. consumers, the near-term meaning is practical: the same assistants that write code and draft documents are increasingly used inside labs to write training code, run experiments and review results. According to IBM Think, AI already assists with coding, experiments and other parts of AI development, creating feedback loops whose limits are now under active study.

What recursive self-improvement actually means

Definitions differ by company, an important source of confusion. According to The Associated Press, some labs use the term for any case where AI output feeds back into model improvement, while others reserve it for systems that upgrade themselves with little or no human direction.

The stronger definition describes a cycle in which a system designs a more capable successor, which then designs another successor. Anthony Aguirre, president and chief executive of the nonprofit Future of Life Institute and a physics professor at the University of California, Santa Cruz, told The Associated Press that autonomous improvement would proceed through successive generations of systems.

Aguirre emphasized speed as the reason the cycle draws attention. He said machine-driven research runs far faster than human-led work, which could compress development timelines if autonomy increases.

Cornell computer-science assistant professor John Thickstun, who studies techniques for steering model behavior, offered a more restrained interpretation. According to The Associated Press, he said a limited form of the process has existed for years because developers routinely use one generation of models to write software used to train the next.

That distinction matters for shoppers and workers evaluating product claims. Assisted coding is already common in consumer developer tools. Self-directed redesign of a frontier model without human review is a different and more demanding step that labs say has not been demonstrated.

Why labs say bigger jumps may be closer

Earlier experiments with models training other models produced only modest gains. According to The Associated Press, Thickstun said work by researchers including OpenAI co-founder Andrej Karpathy yielded small improvements rather than major creative advances.

Current optimism rests on agents that can carry out multi-step engineering work. According to IBM Think, Michael Littman, University Professor of Computer Science and Associate Provost for Artificial Intelligence at Brown University, identified agentic programming systems as an important recent development because they divide projects into stages, generate code, execute programs and revise the result.

Anthropic chief executive Dario Amodei argued this month that AI's expanding role in creating next-generation systems was accelerating overall progress, according to IBM Think. OpenAI chief scientist Jakub Pachocki likewise wrote that progress in automated AI research could eventually point toward recursive self-improvement.

Gabe Goodhart, chief architect of AI Foundations at IBM, described the present dynamic as a flywheel in which stronger models make it easier to build still stronger models, according to IBM Think. He added that people still evaluate many model suggestions before those suggestions are used in another training round.

In other words, the loop is tightening, but humans remain reviewers and decision-makers at key points.

OpenAI, xAI and Microsoft outline different timelines

OpenAI's public roadmap separates near-term assistance from longer-term autonomy. According to The Associated Press, the company said this month that it had built an automated research intern able to handle defined research jobs under human direction, including work that would occupy a skilled researcher for a few days.

The company said it aims to develop an automated researcher by March 2028, a future planned milestone rather than a completed achievement. According to The Associated Press, OpenAI also said recursive self-improvement could help align model behavior with human values, while cautioning that rapid self-improvement is not necessarily a result worth pursuing.

OpenAI said decisions about whether and how to continue should depend on maintaining human control and on informed democratic consideration of benefits and risks, according to The Associated Press.

Elon Musk described a faster trajectory for xAI's Grok series. He said in March that people were gradually becoming less involved in improving the models and that each successive system was constructed with help from its predecessor, while adding that the workflow was not yet completely automated, according to The Associated Press. Musk said full automation might arrive by the end of this year, but no later than 2027, a forecast rather than an accomplished event.

Microsoft is emphasizing a different design philosophy. Mustafa Suleyman, chief executive of Microsoft AI, said the company seeks humanist superintelligence, meaning powerful capabilities that serve people within defined boundaries, according to The Associated Press. In a 2025 essay, Suleyman said that approach would avoid an unbounded system with extensive autonomy in favor of carefully calibrated behavior.

A small experiment shows narrow self-tuning, not an intelligence explosion

Beyond the largest labs, startup Weco AI reported a contained example in July. According to IBM Think, its AIDE² system changed portions of the software framework that governed another automated research agent.

Weco classified the outcome as Level 1 recursive self-improvement under its own framework, meaning a measurable net gain rather than proof of a self-sustaining surge in intelligence. According to IBM Think, one version of the research agent performed better on tasks outside its tuning set, which the company interpreted as broader improvement rather than simple benchmark memorization.

The history of feedback-driven learning is much older. According to IBM Think, Littman noted that researchers have studied improving systems for decades, including reinforcement learning, which rewards actions that produce better outcomes. A more ambitious account of machine-led successor design appeared in Nick Bostrom's 2014 book Superintelligence.

Littman remains unconvinced that present programming agents necessarily lead to full self-redesign. He said it was unclear whether the idea that each new system will necessarily become better at designing AI is logically sound, let alone imminent, according to IBM Think.

For consumers, the practical takeaway is that isolated demonstrations of automatic tuning do not establish that a phone assistant or chatbot can independently rewrite its own foundation model safely.

Safety gaps that worry researchers before full autonomy

Even without complete autonomy, researchers say expanding machine involvement can magnify errors. Nathalie Baracaldo, manager of AI Security and Privacy Solutions and senior research scientist at IBM, told IBM Think that major gaps remain before agents can be trusted to extend themselves without human intervention.

One concern is reward hacking, in which a system finds an unintended method to raise its score without completing the intended job. According to IBM Think, Baracaldo said such behavior occurs when a model maximizes a poorly specified reward while missing the designer's actual objective.

A self-improving setup could worsen that failure if each cycle reinforces the shortcut. According to IBM Think, Baracaldo warned that repeated iterations could push the system toward unsafe conditions and possible real-world harm.

Goodhart pointed to two misuse paths: people deliberately using capable models for harmful ends, and agents operating without oversight and taking actions their operators did not intend, according to IBM Think. He said recursive self-improvement could strengthen both dangers, although additional factors would also be needed for severe harm.

Infrastructure provides one remaining lever of control. According to IBM Think, Goodhart noted that operating the largest models requires substantial computing resources, allowing operators to discontinue access to the running system.

That economic reality does not resolve questions about evaluation, security testing or incident reporting. It does explain why current policy discussion focuses on third-party assessment, transparency about machine contributions and preserving meaningful human review as AI assistance grows.

What to watch next

Three near-term signals will show whether the trend is accelerating responsibly. The first is whether other labs publish the share of research performed by their own models, creating comparable figures to Anthropic's 26% disclosure.

The second is OpenAI's researcher timeline. Its March 2028 target for an automated researcher provides a public date against which outsiders can compare demonstrated capabilities.

The third is xAI's forecast of full automation by the end of 2026 or, at the latest, 2027. According to The Associated Press, Musk presented that schedule as a possibility, not as completed work.

None of those developments would by itself prove that machines can independently generate ever-smarter successors. They would indicate how much routine research labor is shifting from people to software — the concrete change behind this month's warnings and the reason regulators, scientists and consumers are watching measurement and oversight more closely.

Sources referenced