Photo: ZN
Anthropic has warned investors that it may soon reach a stage of recursive self-improvement (RSI), in which artificial intelligence uses itself to develop more powerful models, potentially accelerating technological progress.
However, the company cautioned that such advances could outpace humans’ ability to assess risks and ensure AI systems remain safe.
The warning appeared in an updated S-1A registration statement that Anthropic filed with the US Securities and Exchange Commission (SEC) on November 2, TechCrunch reported.
For critics of AI development, the prospect of each new generation of models improving faster than the previous one poses a serious threat. They fear that systems could become so powerful that humans lose control over their behavior, while the AI may fail to account for the harm it could cause in pursuing its objectives.
Supporters, by contrast, see RSI as an opportunity to boost productivity, accelerate scientific discoveries and create tools capable of improving themselves. Recursive self-improvement is therefore viewed both as a potential technological breakthrough and as a risk requiring oversight.
Anthropic acknowledges both sides. The company says it already uses AI in its own research and development, but warns that doing so could introduce additional safety risks.
“Using AI to build more powerful AI could also pose significant safety risks, especially if it sharply accelerates the pace of progress,” the filing states.
Anthropic said these risks are addressed by its Responsible Scaling Policy, which sets out how the company plans to monitor and mitigate potential threats as it approaches recursive self-improvement.
When the term RSI gained prominence in early 2026, many questioned whether the scenario could be clearly defined, measured or demonstrated in practice. Anthropic’s updated filing provides more specific warnings.
The company acknowledged that it had previously used AI to help develop new models without recognizing at the time the scale of the resulting jump in capabilities.
“We have failed in the past, and may fail in the future, to detect significant acceleration in our rate of progress until after it has occurred,” Anthropic said.
The filing also says the company’s internal assessments and forecasts suggest it could rapidly approach a threshold it defines as “AI-enabled, significantly accelerated progress.” This is a category in Anthropic’s own Responsible Scaling Policy.
The company stressed, however, that none of its current models has crossed that threshold.
The developer of Claude also said it does not plan to give AI full control over the development of future generations of models in the near term.
“We do not expect to be able to safely use Claude for fully autonomous design and development of next-generation models at the frontier without meaningful human oversight or understanding, if and when this becomes technically possible,” the company said.
Anthropic is considering updating its Responsible Scaling Policy to ensure human oversight remains in place during the development of its most powerful models.
However, the company's claims may invite scrutiny. Skeptics point out that an S-1 registration statement is not only intended to disclose information to investors but can also serve a marketing purpose. They suggest that warnings about being close to developing AI capable of designing new models independently could also be intended to highlight Anthropic’s technological capabilities.
Anthropic previously predicted that, by the end of 2026, compact AI models could autonomously replicate themselves through vulnerabilities in everyday consumer devices, without relying on powerful data centers. Researchers have experimentally demonstrated that a network of inexpensive, ordinary TV set-top boxes can support the operation of an AI model.
