Recursive self-improvement (RSI) is a process in which an AI system helps build a more capable successor, and that successor then helps build the next one. On November 2, 2026, Anthropic said its future models may soon reach a point where this process speeds up AI development sharply. The statement appears in an amended S-1 registration filing with the US Securities and Exchange Commission (SEC).
This article explains what the filing says, how Anthropic defines its threshold, and what the disclosure means for AI safety and human oversight.
What Did Anthropic Say About Recursive Self-Improvement in its SEC Filing?
Anthropic stated that its internal estimates and its expectations for future models, not its current models, suggest it may soon meet the Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI,” according to Gizmodo’s report on the filing.
The company also stated that none of its current models has crossed that threshold. El-Balad’s summary of the TechCrunch report says the filing links the risk to Anthropic’s growing use of AI in research and development, which could push progress ahead of its ability to build matching safety, alignment and security measures.
The filing adds that Anthropic is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.
What is Recursive Self-Improvement in AI?
Recursive self-improvement is a feedback loop in which AI systems contribute to the research, coding, experimentation and optimisation that produce the next generation of AI systems. Each generation could make the following one faster or cheaper to build.
The term does not mean an AI system can freely redesign itself or operate outside human control. It covers a range of capabilities, and how reliably a model can improve its successors remains an open technical question.
What is Anthropic’s Threshold For “Dramatically Accelerated Progress Attributable to AI”?
Anthropic uses this phrase instead of the term RSI. Gizmodo’s reading of the Responsible Scaling Policy is that the threshold is concrete: the baseline rate of AI progress doubles, and the doubling is plausibly attributable to AI.
That detail turns a vague fear into something measurable in principle. It also shows why measurement is difficult. The filing acknowledges that researchers may not recognise a major acceleration until after it has happened.
What is the Difference Between AI-Assisted Research and Recursive Self-Improvement?
AI-assisted research uses AI as a tool under human direction. Recursive self-improvement means the AI helps design and build its own successor.
| Factor | AI-assisted research | Recursive self-improvement |
| Role of the AI | Tool that speeds up human work | Contributor that helps design and build its successor |
| Human involvement | Humans direct each major step | Humans supervise a loop they may struggle to follow |
| Current status at Anthropic | Already in use | Not declared as reached |
| Main risk | Ordinary errors and overreliance | Progress outpacing safety testing and governance |
Anthropic’s warning concerns how close future models may come to the threshold. The company has not announced that recursive self-improvement has been achieved.
Why Does Anthropic’s RSI Warning Matter for AI Safety and the AI Industry?
The warning matters because frontier AI companies already use their own models to write code, run research and improve internal processes. Anthropic has said publicly that Claude now writes more than 80% of the code merged into its codebase, as Fortune reported in June. That figure shows the starting point for the loop the filing describes.
Faster development brings clear commercial benefits in software, scientific research and cybersecurity. It also gives safety teams less time to study new capabilities, investigate unexpected behaviour and set restrictions before release.
The filing is also a legal document. An S-1/A exists to disclose material business risks to prospective investors. By placing AI-driven research acceleration in that section, Anthropic is telling investors that the technology behind its growth could produce risks that are hard to predict or manage.
What Questions Does Anthropic’s Filing Leave Unanswered?
Three questions remain open:
- What evidence supports the estimate? The filing does not say what specific internal data led Anthropic to conclude that future models may meet its threshold.
- When will the policy change? Anthropic is considering updates to the Responsible Scaling Policy but has not given a date.
- How will anyone verify the threshold? If acceleration is difficult to see in real time, outside parties have limited means to check the company’s claims.
Author’s Note:
I read the filing as one piece of a pattern, not as a standalone alarm. In February, Time reported that Anthropic had removed the central commitment of its Responsible Scaling Policy, which was a pledge not to train a model unless safety measures were already adequate, as Fortune summarised. In June, the company published a paper calling for a coordinated slowdown among frontier labs, as The Next Web’s timeline records. Now the company has put a possible RSI threshold into a securities filing.
Taken together, these steps show a company that treats AI-accelerated research as a near-term governance problem. They also show a company whose own safety rules have loosened while its warnings have grown louder. Both facts deserve equal weight.
The practical test is whether Anthropic publishes the revised policy, the evidence behind its estimate, and a method that outside researchers can use to check the threshold. Until then, the warning is credible as a signal and incomplete as a safeguard.
The next challenge for the industry is to show that human oversight, safety evaluations and governance can keep pace with the capabilities AI companies are building.
