Where The Evidence Stands On Recursive Self-Improvement In AI
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Where The Evidence Stands On Recursive Self-Improvement In AI on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Recent evidence shows AI systems assisting with AI research, and an independent evaluator says Anthropic’s progress on well-specified tasks may be running at about twice its earlier rate. The evaluator attributes most of that acceleration to AI assistance, not fully autonomous work; no documented system has independently designed and built its successor. Two congressional bills proposing restrictions remain in committee.

AI is helping researchers develop AI, but the evidence available as of November 16 does not document a system independently designing and building its own successor. The distinction has drawn fresh attention after podcaster Dwarkesh Patel published a viral essay on November 13 and congressional proposals to restrict AI development remain pending.

The clearest recent evidence comes from METR, an independent evaluator, which reviewed internal Anthropic data from August 12 to October 29, 2026. It concluded that Anthropic was likely improving model capabilities on well-specified tasks at about twice the rate seen from March 2024 to October 2025. METR attached high uncertainty to that estimate, saying the actual acceleration could plausibly be closer to 1.3 times or three times.

METR said most of the speed-up was likely due to AI assistance rather than fully autonomous contributions. It also said AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. The evaluator noted that Anthropic redacted much of the evidence it considered most informative, limiting what outsiders can verify.

Anthropic’s own August figures describe a human-supervised process: Claude “leads” 26% of the company’s AI research and development work, while more than 90% of that work involves AI at the “collaborates” level or above. Anthropic said no measured subset of its AI research and development ran fully autonomously. The company defines the more complete form of recursive self-improvement as an AI system autonomously designing and developing its successor, and says that has not been reached.

At a glance
reportWhen: Developing as of November 16, 2026; the…
The developmentA viral essay has renewed debate over recursive self-improvement as published company data and an independent evaluation show AI-assisted research, but not autonomous AI systems building their own successors.

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

AI collaborates or moreabove 90%
AL3 and up
AI leads, a human supervises26%
AL4
Fully autonomousno measured part
AL5 · the closed loop

Anthropic: “We are not there yet.”

The bill’s line: “the capacity to automate or greatly accelerate” AI R&DS.5493 / H.R.10538 list it as a “superintelligence precursor”. How much is “greatly” is not defined.
  1. 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.
  2. AI leads about a quarterIn 26%, Claude completes most of a task from a high-level prompt while a human supervises.
  3. No closed loopNo measured part runs fully autonomously. That is the step “recursive self-improvement” usually means.
  4. 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

Why Human Oversight Changes the Claim

The distinction matters because AI-assisted research and autonomous self-improvement are different levels of capability. A system that helps engineers write code, test ideas or complete research tasks may speed up development while people still set goals, make decisions and oversee the work. That is evidence of AI contributing to AI progress, but it does not by itself show that an AI can take control of the full research-and-development cycle.

For readers, businesses and policymakers, the practical question is not just whether AI contributes to research. It is which tasks it performs, how reliably it does them, and where human judgment remains necessary. The published evidence suggests companies are using AI more extensively in research. The available evaluation does not establish that systems can independently make the most important training decisions or develop a successor without human involvement.

Those differences also shape the debate over regulation. Legislation that covers AI systems capable of accelerating AI research could apply well before a system can autonomously build another model. The bills described in the source material would reach beyond the narrowest definition of recursive self-improvement, making their exact scope relevant to developers and users alike.

De'Longhi La Specialista Touch Semi-Automatic Espresso Machine with Grinder

De'Longhi La Specialista Touch Semi-Automatic Espresso Machine with Grinder

De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother

De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother

As an affiliate, we earn on qualifying purchases.

How the Evidence and Bills Fit

Recursive self-improvement, or RSI, usually refers to a feedback loop in which an AI helps create a more capable AI, which can then help create the next one. But discussion often combines several stages: AI tools assisting people, AI leading parts of research under supervision, and AI carrying out the entire process on its own. Anthropic uses an automation scale that runs from no AI involvement to fully autonomous operation; its reported Claude figures place some work at the “lead” level, not the fully autonomous endpoint.

OpenAI has also reported progress toward automating research tasks. According to The Decoder, which reviewed an OpenAI blog post that was unavailable to the source author, the company said it had reached its “automated research intern” goal and aims for a full automated AI researcher by March 2028. The Decoder reported that tasks shorter than 15 minutes succeeded 86% of the time without intervention, while more than half of successful tasks lasting four to eight human hours needed at least one human step. These are company-reported measurements, not independently reviewed results, according to The Decoder.

The political debate concerns 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 on September 24. The bills would pause training, modification and fine-tuning of large AI systems until a proposed Department of Artificial Intelligence establishes rules. Their text includes systems able to automate or greatly accelerate AI research and development, including through recursive self-improvement. The latest status cited in the source shows both bills referred to committee; the House bill had 12 Democratic cosponsors. A separate proposal from Rep. Ro Khanna was reported as a summary, with no bill text located by the source.

Patel’s November 13 essay, “The avalanche that hit math is coming for AI development,” draws on METR’s assessment and figures published by Anthropic and OpenAI. Its further predictions are commentary and forecasts, not new experimental findings. The renewed political attention also follows the Associated Press calling both chambers of Congress for Democrats on November 3, which supporters of restrictions have cited as a possible opening in the next Congress.

“Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.”

— METR

What the Measurements Cannot Show

The evidence does not establish autonomous recursive self-improvement, but it also offers only a partial view of how AI contributes to research inside leading companies. METR’s assessment relied on internal Anthropic data, and it said Anthropic redacted a majority of the evidence it found most informative. Its estimate of acceleration carries a wide uncertainty range and applies to well-specified tasks, not every part of AI development.

OpenAI’s reported figures are also company measurements, as described by The Decoder, rather than independently reviewed results. The available material does not provide a shared measurement standard for comparing the two companies, nor does it settle how much AI assistance has changed overall development timelines. It remains unclear how quickly AI systems may move from completing supervised tasks to making significant research decisions without human intervention.

The status of the legislative proposals could change, but the cited records show both Sanders-Casar bills referred to committee. The source found no bill text for Khanna’s separate proposal. Nothing in the material indicates that these proposals have become law or changed requirements for businesses or users.

The Next Evidence and Legislative Steps

The next useful evidence would include independent evaluations with fuller access to underlying data, clearer descriptions of which research tasks AI completes, and measurements that separate assistance from autonomous decision-making. Anthropic and OpenAI may publish further results, but company claims should be distinguished from outside evaluations and from forecasts about future capability.

On the policy side, the Sanders-Casar bills would need action in committee before advancing. Their current status does not impose a pause on AI development. Supporters may pursue the proposals in the next Congress, while details of Khanna’s proposal remain unavailable in the source material. Until new evidence or legislative action emerges, the grounded conclusion is that AI is already involved in parts of AI research, while a system independently creating its own successor has not been documented.

Key Questions

Has recursive self-improvement started?

AI-assisted AI research is documented, including systems that lead some work under human supervision. Fully autonomous recursive self-improvement, defined as an AI designing and developing its own successor, has not been documented in the cited evidence.

What did METR find about Anthropic’s development speed?

METR estimated that Anthropic was improving model capabilities on well-specified tasks at about twice its earlier rate, comparing October 2026 with March 2024 to October 2025. It marked the estimate as highly uncertain and said most of the speed-up likely came from AI assistance.

Can AI complete research tasks without human help?

Some short tasks reportedly succeeded without intervention, according to OpenAI figures reported by The Decoder. For successful tasks lasting four to eight human hours, however, more than half needed at least one human step. These were company-reported measurements, not independent verification.

Would the proposed U.S. bills ban AI development now?

No. The Sanders-Casar bills were referred to committee in the latest status cited, and the source says they have not moved. They are proposals, not current restrictions on businesses or users.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

October 2026: What an Anthropic IPO Actually Unlocks

Anthropic’s planned October 2026 IPO, valued at up to $900B, is a structural event that will reshape AI industry dynamics and market expectations.

Muse Wonder Full Joins The AI Frontier: What To Know

Artificial Analysis scores Muse Wonder Full at 56, but the API model costs $4.52 per benchmark task and trails cheaper rivals on price.

Abu Dhabi’s AI-Native Government Goal Could Redefine Public-Sector Work

Abu Dhabi and OpenAI launched a food-import compliance pilot as part of four agreements, including a plan to make selected rules machine-readable.

The Forecast Is the Plan.

Major AI labs publicly commit to automating AI R&D, with OpenAI targeting a research intern by September 2026. This signals a shift in AI development strategies.