🔍 Read the full analysis: What Counts As Recursive Self-Improvement In AI—and Could It Be Banned? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Evidence cited by Anthropic and independent evaluator METR shows AI tools helping researchers develop AI, with people still setting goals and overseeing work. A bill introduced in Congress would pause training and modification of large AI systems, including through recursive self-improvement, but it has not advanced beyond committee referral.
AI systems are helping researchers develop AI, but available evidence does not show a system autonomously designing and building its own successor, according to research cited by Anthropic and independent evaluator METR. The distinction is central to a debate revived this week by podcaster Dwarkesh Patel’s essay and a congressional bill that would pause some AI development, including work involving recursive self-improvement.
Recursive self-improvement, or RSI, commonly describes a feedback loop in which an AI system helps create a more capable system, which then helps create the next one. Anthropic defines the more advanced endpoint as an AI capable of fully autonomously designing and developing its own successor. The company says that has not happened: its August 2026 figures show Claude leading 26% of its AI research and development work, while no measured subset operates fully autonomously.
Anthropic also reports that more than 90% of its AI research and development work involves AI at the level of collaboration or above. It uses an automation scale from Epoch AI, running from no AI involvement to fully autonomous operation. On that scale, “leads” means AI completes most of a task from a high-level prompt while a human supervises. These company figures describe Anthropic’s own work, not the entire AI industry.
METR, a nonprofit evaluator, reviewed internal Anthropic data covering August 12 to October 29, 2026. It estimated that Anthropic’s progress on well-specified tasks was running at about twice the rate of the period from March 2024 to October 2025, but emphasized high uncertainty: its estimate could be closer to 1.3 times or three times. METR attributed most of the apparent acceleration to AI assistance, while saying AI had not made highly significant, fully autonomous contributions to the most important parts of model training. It also said Anthropic had redacted much of the evidence it considered most informative.
The policy proposal is the Ban Artificial Superintelligence Act of 2026, introduced by Sen. Bernie Sanders as S.5493 on September 23 and Rep. Greg Casar as H.R.10538 the following day. The bills would pause training, modification and fine-tuning of large AI systems until a new Department of Artificial Intelligence set rules. Their scope includes recursive self-improvement and systems with the capacity to automate or greatly accelerate AI research and development. The latest status described in the source material shows both bills referred to committee; the House bill had 12 Democratic cosponsors.
Has it started? · explainer
AI already does much of the work of building AI. Not alone.
How far AI research and development is automated at one lab, on a scale from AL0 (no AI) to AL5 (no human in the loop), and where the proposed ban would draw its line.
Anthropic’s AI R&D work, August 2026
Automation levels from Epoch AI’s scale, as measured by Anthropic
Anthropic: “We are not there yet.”
- AI collaborates on most of the workAbove 90% of the work is at or above “AI collaborates”, where AI does large chunks under close human direction.
- AI leads about a quarterIn 26%, Claude completes most of a task from a high-level prompt while a human supervises.
- No closed loopNo measured part runs fully autonomously. That is the step “recursive self-improvement” usually means.
- Where would a ban start?The bill names “greatly accelerate” as a warning sign, so its reach depends on a measurement nobody yet makes in a standard way.
Sources: Anthropic, “Measurements for understanding the pace of AI development inside frontier labs” and “When AI builds itself” (figures as of August 2026, Anthropic’s own measurements); Ban Artificial Superintelligence Act of 2026, S.5493 and H.R.10538, sec. 3(6)(A). Bars show shares of Anthropic’s AI R&D work.
Graphic: Powered by Thorsten Meyer AI
The Line Between Assistance and Autonomy
The policy debate turns on what counts as self-improvement. If the term includes AI helping human researchers, it covers work already reported by companies. If it means a system independently building its successor, the cited evidence does not establish that this has occurred. The distinction could determine whether a proposed restriction reaches ordinary AI-assisted research or only more autonomous development.
The bill’s language also reaches beyond a narrow definition of RSI: it would cover systems capable of automating or greatly accelerating AI research and development. That could make the boundary relevant to labs, researchers and institutions well before any system independently designs its successor. The source material says the proposals have not advanced, so they do not currently change what businesses or users can do.
METR’s estimated pace of progress is another reason the issue has drawn attention, but it is not proof that a self-improving loop is underway. The estimate concerns capability gains on well-specified tasks at one company, carries a broad uncertainty range and relies on internal evidence that was substantially redacted. Readers should distinguish that finding from Patel’s forecast about what might happen next.
Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

DJI Osmo 360 Road Cycling First-Person Combo Action Camera

Bose SoundLink Max Portable Bluetooth Speaker, IP67 Waterproof and Dustproof, Up to 20 Hours of Battery Life, Black
As an affiliate, we earn on qualifying purchases.
Evidence Behind the Renewed Debate
Patel published “The avalanche that hit math is coming for AI development” on November 13. The essay and its narrated version drew attention to the possibility that AI could speed up the development of more capable AI. The source describes Patel’s argument as commentary based on existing assessments and figures from Anthropic and OpenAI, rather than new evidence that autonomous self-improvement has begun.
OpenAI said in a September blog post that it had reached its goal of an “automated research intern,” according to its own measurements, and set a goal of a full automated AI researcher by March 2028. The source material reports these details through The Decoder, which read the post after the original site blocked the article’s download. It says tasks under 15 minutes succeeded 86% of the time without intervention, while more than half of successful tasks lasting four to eight human hours required at least one human step. OpenAI says people still set research priorities and make decisions about scaling, pauses and deployment. The company’s figures are internal and, according to The Decoder, lack independent review.
Political attention has also shifted. The source says supporters of a ban see a possible opening in the next Congress after Democrats won both chambers in the November 3 elections, as reported by the Associated Press. Rep. Ro Khanna has also proposed a “Human Control Over AI Act,” but the source found only a September summary seen by CNBC, not bill text. The scope of that proposal cannot be verified from the available material.
““We are not there yet, and recursive self-improvement is not inevitable.””
— Anthropic, in its essay “When AI builds itself”
Limits of the Available Evidence
No fully autonomous successor-building system is documented in the evidence described here. That does not establish that no such work exists elsewhere; the source material presents company reports and one evaluator’s review, not a comprehensive audit of the field.
METR’s estimate has substantial uncertainty, and it said much of the evidence it found most informative had been redacted by Anthropic. OpenAI’s figures also come from inside the company and were not independently reviewed, according to The Decoder. How much AI assistance causes faster development across the industry, rather than at these particular companies, remains unclear.
The exact reach of the proposed federal pause would depend on how its terms were interpreted and on any rules made by the proposed department. Neither bill has advanced beyond committee referral in the status provided. The source does not establish whether the proposals will receive hearings or votes, and it provides no text for Khanna’s separate proposal.
Committee Status and New Evidence
The immediate legislative milestone is whether either Sanders’s or Casar’s bill moves out of committee. Until then, the proposals remain bills, not restrictions in force. The source material does not give a scheduled hearing, vote or implementation date.
For the technical debate, the next useful evidence would include further independent evaluations, clearer disclosure of how AI contributes to research, and results showing whether systems can complete longer, more consequential development tasks without human intervention. OpenAI’s stated target of a full automated AI researcher by March 2028 is a company goal, not a confirmed outcome or independently validated forecast.
Key Questions
Has recursive self-improvement started?
AI-assisted AI research is documented in company reports. The cited evidence does not show a system independently designing and building its own successor, which Anthropic identifies as the fully autonomous endpoint.
What does the proposed ban cover?
The Ban Artificial Superintelligence Act of 2026 would pause training, modification and fine-tuning of large AI systems until a proposed federal department set rules. Its text includes recursive self-improvement and systems able to automate or greatly accelerate AI research and development.
Is the bill already in force?
No. The source material says the Senate and House bills were referred to committee and had not advanced. They do not currently impose a pause on AI development.
How reliable is the claim that AI has sped up AI research?
METR estimated that Anthropic’s progress on well-specified tasks was about twice as fast as during an earlier period, but gave a wide uncertainty range and said much of the acceleration was likely due to AI assistance. Its review relied on internal data, some of which Anthropic redacted.
Source: ThorstenMeyerAI.com
