Recursive self-improvement showed up in Anthropic’s SEC filing and an independent audit. Here is what it means, and why you need to know about it.
A new AI related term has entered the chat: RSI, short for recursive self-improvement. Within one week, an AI company warned about it in an investor filing, an independent group published an audit of it, and Members of Congress are calling for a ban. Most people in Washington have never heard of it. They should.
What It Is
RSI is what happens when AI gets good enough at building AI that the work starts to speed itself up. Picture a company whose engineers use its own product to design the next version, and the product keeps getting better at the engineers' jobs. Each generation helps build the next.
People use the term two ways, which causes a lot of confusion. In the broad sense, any AI that helps build better AI counts, even if people review and direct every step. That's mostly what's happening now. In the strict sense, RSI means a "closed loop": AI does the research, checks its own work, and builds the next version with no human review. That's the version most call for a ban focus on, and where the most serious safety concerns lie. In between, people no longer review every step but still audit the results and decide what ships.
What Just Happened
In Anthropic’s recent IPO filing with the SEC, listed in its risk factors it said it has in the past failed to recognize a big acceleration in its own progress until after it happened. It also said it may soon meet its own definition of "dramatically accelerated progress attributable to AI." It added that none of its current models have crossed that threshold, and that it doesn't expect to safely let AI design its successors without meaningful human supervision.
Three days later, the nonprofit METR published an independent assessment after spending 11 weeks inside Anthropic. It estimated that the company's models are improving about 1.9 times faster than before, with a likely range of 1.5 to 2.8 times, and that most of the speedup comes from Anthropic's own use of AI.
Congress Is Already Moving
Last September, U.S. Sen. Bernie Sanders (I-VT) and U.S. Rep. Greg Casar (D-TX) introduced the Ban Artificial Superintelligence Act, which would ban superintelligence, regulate systems with precursor capabilities including RSI, and create a Cabinet-level Department of AI. On November 6, New York Assembly member Alex Bores wrote, "Congress should ban RSI, this year if possible," and Casar replied, "Ban RSI." One summary of the debate says supporters of a narrower RSI ban cross party lines, from Sanders to Steve Bannon.
We're not endorsing or opposing any proposal. But understanding the debate matters.
Where China Stands
Chinese labs are working on this too. Companies including Zhipu, DeepSeek, Alibaba, ByteDance, and Moonshot AI have published or announced work on using AI to speed up their own research, mostly on the machinery around models, such as training pipelines, and not on the models themselves. Chinese officials have also named the risk: in September, the country's AI standards body warned of a "self-accelerating trend" that includes recursive self-improvement. Analysts disagree on how close China is, and Chinese firms remain short of advanced chips. The September Trump-Xi summit set up a U.S.-China AI dialogue, the natural place for any joint limits. A ban in one country wouldn't bind the other.
What Washington Should Know
RSI isn't one thing. The policy question isn't whether AI helps build AI. It's how much human oversight stays in the loop, and how anyone would know.
It's hard to define in law. A ban requires picking a threshold. AI is already writing much of the code at leading labs, and the line between assistance and autonomy is blurry.
The evidence is thin and mostly self-reported. Washington currently learns about this from company filings and a partly redacted audit.
Ban versus race is a false choice. The country can want the benefits of faster AI and still ask for verification.
Questions to Ask
Who independently checks a company's claims about its own AI research?
What would count as a warning sign that AI is speeding up its own development faster than people can check it, and who decides?
What does the government need access to, and what can stay confidential?
How do we keep building, and keep humans in charge?
Who do you reach out to if something goes wrong (or is uncontrollable)?
What We Don't Know
At this time we don’t know if any lab is close to a closed loop. We also don’t know whether the acceleration will last or plateau, or whether a ban could be written in a way that works.
Sources: SEC- anthropicpbc,Yahoo Finance, METR, METR on X, Ban Artificial Superintelligence Act, x.com/RepCasar, MindStudio, Pandaily, SemiAnalysis, https://www.cnbc.com/2026/09/24/trump-xi-china-summit-ai-export-control-nvidia-huawei-alibaba-.html, https://www.cnbc.com/2026/09/25/the-tech-download-trump-xi-ai-talks.html




