Recursive self-improvement moves from AI theory into near-term lab roadmaps

Recursive self-improvement is moving from a distant AI thought experiment into the near-term planning of frontier laboratories. AP reported that leading developers now say AI systems are approaching the point where they can help improve themselves and contribute to building more advanced successor models. The idea, often shortened to RSI, has long been a central concern in AI safety debates because it could accelerate scientific progress while also making systems harder for humans to understand, monitor or restrain.

The clearest recent evidence comes from Anthropic’s disclosure about Claude’s role inside its own research and development work. Anthropic said Claude is now leading about 26 percent of its model research and development tasks under human supervision. The company’s description does not mean Claude is independently designing the next generation of AI, but it does show that models are becoming part of the machinery used to build future models. In a frontier lab, even partial automation of coding, evaluation, data work and experiment design can compound quickly.

OpenAI has also described progress toward automated AI research. AP reported that OpenAI announced an automated “research intern” capable of carrying out well-defined research tasks under human direction, including work that could take a skilled researcher days. The company has said it is working toward a more autonomous AI researcher by March 2028, while also warning that it does not yet know how to safely reach aligned, full RSI. Elon Musk has described xAI’s Grok development process as moving toward less human involvement, with a target for greater autonomy by 2027.

The disagreement is not about whether AI can help AI researchers. That is already happening. The harder question is whether the feedback loop can become strong enough that each new system helps build a more capable next system, which then improves the process again. Supporters argue that such systems could speed breakthroughs in medicine, materials science, climate modeling and safety research itself. A tireless automated researcher could test hypotheses, inspect code, run experiments and find errors at a scale humans cannot match.

Critics worry about the same acceleration for a different reason. If models start improving the systems that train or evaluate them, companies may lose the ability to predict where capabilities are headed. Safety researchers often describe the runaway scenario as one where improvement speed increases faster than oversight, leaving humans with less time to test, understand or intervene. Even more grounded versions of the risk matter: automated research agents could introduce subtle errors, optimize for the wrong metric or create tools whose behavior is difficult to audit.

Microsoft has framed its own approach as “humanist superintelligence,” emphasizing systems that remain carefully limited and in service of people rather than unbounded autonomous entities. OpenAI has similarly said that decisions about rapid RSI should depend on preserving human control and democratic choices about benefits and risks. Anthropic has encouraged other labs to publish comparable metrics about how much AI contributes to model development.

The new AP reporting shows why RSI is becoming a practical governance issue rather than a speculative philosophy topic. Once AI becomes part of the research pipeline, companies need records of what agents did, which tasks humans reviewed, when safety systems blocked actions and how model-assisted research changes the risk profile of future releases. Recursive self-improvement may still be incomplete, but the first loops are visible enough that labs, regulators and the public can no longer treat it as a distant scenario.