Anthropic, the company behind Claude, has warned that its models could soon reach a point where artificial intelligence plays a central role in building its own successor. The warning, in an updated filing to the United States Securities and Exchange Commission on November 2, does not claim that moment has arrived. It says the company’s internal estimates suggest a major threshold, known as recursive self-improvement, may soon be within reach.
Recursive self-improvement describes an AI system good enough at research and engineering to help create a more capable version of itself, which then helps create an even more capable version. No single step need be a dramatic leap. The concern, and the promise, lies in compounding: small improvements, repeated quickly, could produce a much faster pace of change than planned. Today, researchers already use AI tools to write and check code, analyse data and test models, but major decisions about design, training and release still rest with people.
The filing also makes an uncomfortable admission: Anthropic says it has previously failed to recognise significant increases in its own development speed until after they had happened. If progress can accelerate quietly, only becoming obvious in hindsight, then the systems meant to test, monitor and restrain new models may find themselves playing catch-up. A formal warning of this kind invites regulators, researchers and the public to ask what safeguards should already be in place, rather than waiting for a breakthrough to force the question.
AI Development Could Soon Speed Up Sharply
Anthropic says it is already using artificial intelligence to accelerate its research and development. Across the industry, AI assistance is now routine in software work, and the most capable models are used to probe, test and refine their successors. Anthropic adds a warning about the pace that could follow: its estimates for future models suggest it could soon meet its own definition of dramatically accelerated development. The wording is measured, speaking of estimates rather than a confirmed result, and of future models rather than today’s systems.
That raises a practical problem for oversight. Safety testing takes time: a model must be probed for unexpected behaviour, checked across many scenarios, and assessed for risks that may only appear at scale. If the model is changing rapidly, or helping to produce its successor before it has been fully understood, those checks become a moving target. That even well-resourced teams with direct access to their systems can underestimate the pace makes the concern harder to dismiss.
The Safety Risks Behind Faster Progress
The safety argument follows directly from that pace. If capabilities improve faster than the systems designed to keep them safe, researchers may struggle to spot unexpected behaviour, evaluate emerging risks, or ensure advanced models keep following human instructions. Testing adequate for one generation may not suffice for the next, especially if it arrives sooner than expected.
Anthropic frames these concerns through its Responsible Scaling Policy, its guidelines for monitoring advanced capabilities and reducing risk. The policy is, in effect, a promise to pause or slow down if warning signs appear, and to add safeguards as models grow more powerful. It will be tested by the very acceleration it describes, since a policy only works if those applying it can see risks clearly and act before they become unmanageable. Anthropic also warns that rapid development could raise the risk of incidents and attract further regulatory scrutiny.
Humans Must Remain in Control
Despite the warning, Anthropic is explicit about its intentions. It says it does not plan to allow Claude to develop new generations of models independently, without human supervision, even if fully automated development becomes technically possible. Meaningful human involvement will remain necessary, it says. It is also considering updates to its Responsible Scaling Policy, through extra testing, restrictions on releases, and other measures to keep capabilities within its ability to manage them safely.
Does This Mean AI Can Already Build Itself?
The short answer, on Anthropic’s own account, is no. Its filing does not establish that Claude has achieved full recursive self-improvement, or that AI can currently develop new generations of models without human help. It warns only that AI-assisted research may soon cross a specific threshold under its safety policy. That threshold is a warning light, not a finish line, and the filing gives no confirmed timeline for fully autonomous capabilities.
The longer answer is where the real debate lies. As AI systems improve at writing code, solving mathematical problems and assisting research, companies will increasingly rely on them to build the next generation of AI. That reliance may grow gradually, until the human role shifts from designing systems directly to supervising systems that design other systems. Whether researchers can maintain effective oversight as that process accelerates remains one of the industry’s biggest unanswered questions. Supporters will point to faster discovery and new medicines, but Anthropic’s filing puts the question on the formal record, from inside a company best placed to know.
