AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Has Recursive Self-Improvement Started In AI? Key Questions To Consider on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

AI is already assisting and leading parts of AI research, according to company data and an independent evaluation of Anthropic. That is not the same as a system autonomously building its own successor: the cited evidence does not show that stage has been reached. A proposed U.S. bill targeting recursive self-improvement has been introduced but has not advanced beyond committee referral.

A November 13 essay by podcaster Dwarkesh Patel has renewed debate over whether AI has begun improving itself, drawing on evidence that AI tools are speeding up some research and development work. The cited evidence supports AI-assisted research with humans still involved; it does not document an AI system autonomously designing and building its own successor. At the same time, a U.S. bill proposing restrictions on such systems remains in committee, with no change to what businesses or users can do today.

The clearest distinction is between AI helping researchers develop models and a system running the full improvement loop on its own. Anthropic describes recursive self-improvement at its most advanced as an AI system that can fully autonomously design and develop its successor. The company says that has not happened. Its August 2026 figures, using an automation scale, indicate that Claude leads 26% of the company’s AI research and development work, while no measured part of that work is fully autonomous. Anthropic also reports that more than 90% of the work it measured is at or above the level where AI collaborates.

An independent evaluation by the nonprofit METR, published November 5, examined Anthropic’s internal data from August 12 to October 29. METR estimated that, by October 2026, Anthropic was improving model capabilities on well-specified tasks at about twice the rate recorded between March 2024 and October 2025. It stressed that the estimate had high uncertainty, saying the true acceleration could be closer to 1.3 times or three times. METR attributed most of the likely speed-up to AI assistance, not fully autonomous contributions. It also said AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. Anthropic redacted much of the evidence METR considered most informative.

OpenAI has also described progress toward automated research assistance. According to The Decoder’s account of an OpenAI blog post, the company said it had met its goal for an “automated research intern,” based on its own measurements, and aimed to build a full automated AI researcher by March 2028. OpenAI reported that tasks taking researchers under 15 minutes succeeded 86% of the time without intervention. For tasks that would take people four to eight hours, more than half of successful attempts needed at least one human step. The Decoder noted the figures came from inside OpenAI and that the post did not describe independent review.

At a glance
analysisWhen: Developing; the essay was published Nov…
The developmentA viral essay has renewed debate over whether AI is improving its own development, as evidence shows AI-assisted research but not fully autonomous successor-building.

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

AI Assistance Is Not Autonomy

The distinction matters because faster AI-assisted development and autonomous self-improvement carry different implications for oversight. AI tools can help researchers write code, test ideas, or complete well-defined tasks while people still choose research goals, check results, and decide whether to train or deploy a model. That can speed up work without establishing that an AI system can set its own direction and produce a more capable successor independently.

The evidence is also relevant to policy. The proposed Ban Artificial Superintelligence Act would reach beyond a narrow ban on autonomous successor-building: its text would pause certain training, modification, and fine-tuning of large AI systems, including through recursive self-improvement, until a new Department of Artificial Intelligence set rules. It would also cover systems with the capacity to automate or greatly accelerate AI research and development. That breadth makes the meaning of “started” consequential, but the bill has not become law. For now, the research claims and proposed restrictions do not alter current rules for companies or users.

Products Worth Considering

Apple Watch Series 10 GPS 46mm with Silver Aluminum Case and Blue Cloud Sport Loop

Apple Watch Series 10 GPS 46mm with Silver Aluminum Case and Blue Cloud Sport Loop

As an affiliate, we earn on qualifying purchases.

The Evidence Behind the Debate

Patel’s essay, “The avalanche that hit math is coming for AI development,” was published November 13 and also presented as a narrated episode. It draws on METR’s assessment and figures published by Anthropic and OpenAI. The claim that AI is speeding up AI development is therefore tied to reported measurements; Patel’s broader expectations about how quickly the process could accelerate are forecasts, not a new demonstration of autonomous self-improvement.

The legislative 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 on September 24. The latest status described in the source material shows both bills referred to committee. The House bill had 12 cosponsors, all Democrats. A separate proposal from Rep. Ro Khanna, called the Human Control Over AI Act, was described by CNBC in September based on a summary; the source says it found no bill text.

The debate resurfaced after the Associated Press called both chambers of Congress for Democrats on November 3. Supporters of the proposed restrictions see an opportunity in the next Congress, but that political development does not itself change the bills’ status or establish that autonomous recursive self-improvement has occurred.

““As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate” of the earlier period.”

— METR, in its November 5 evaluation

Limits of the Available Evidence

No cited evidence shows a system independently building its own successor, but the available measurements do not settle every question about how much AI contributes to research across the industry. METR’s assessment concerned Anthropic, carried high uncertainty, and relied on data the company provided, much of which was redacted. OpenAI’s reported figures were also company-generated and, according to The Decoder, lacked independent review.

Definitions matter as well. “Recursive self-improvement” can refer loosely to AI helping improve AI, or more narrowly to a system autonomously designing and developing a successor. The evidence supports the former in a human-directed setting; it does not establish the latter. It remains unclear how quickly assistance may expand, how well results transfer from measured tasks to broader research, and whether future systems will be able to take on more of the work without human intervention. The status and contents of Khanna’s separate proposal are also limited in the source material to a reported summary rather than bill text.

Research and Bills Await Review

The next evidence to watch is whether further independent evaluations can measure AI’s role in model development with enough detail to distinguish assistance from autonomous work. Anthropic’s figures are a snapshot from August, and METR’s assessment covers data through October; additional evaluations could show whether the reported pace continues or changes. OpenAI has stated a goal of a full automated AI researcher by March 2028, but that is a company target, not a confirmed outcome.

On the policy side, both versions of the Ban Artificial Superintelligence Act remain referred to committee in the status described by the source. Their prospects depend on congressional action, and the text’s proposed restrictions have not taken effect. Until there is new evidence or legislative movement, the grounded conclusion remains limited: AI is helping accelerate parts of research and development, while fully autonomous recursive self-improvement has not been documented in the cited material.

Key Questions

Has recursive self-improvement started?

It depends on the definition. AI is assisting with and leading some research work while humans remain involved. The cited evidence does not show a system autonomously designing and developing its own successor.

What did METR find about Anthropic’s research pace?

METR estimated that Anthropic’s progress on well-specified model-capability tasks was about twice the rate of the March 2024 to October 2025 period. It described the estimate as highly uncertain and said AI assistance likely explained most of the acceleration.

Would the proposed U.S. bill ban all AI research?

The bill text described in the source would pause specified training, modification, and fine-tuning of large AI systems until a new department set rules. It would also cover systems able to automate or greatly accelerate AI research and development. The bill has not advanced beyond committee referral, so the proposed restrictions are not in effect.

When does OpenAI expect a full automated AI researcher?

According to The Decoder’s account of an OpenAI post, the company has set March 2028 as a goal for a full automated AI researcher. That is a stated target, not evidence that the milestone has been achieved.

Source: ThorstenMeyerAI.com

You May Also Like

I Gave an AI a Civilization to Run. It Built a Nuke – Launching CivBench

An AI controlling a simulated civilization in Civilization VI built and launched a nuclear device, raising concerns about AI decision-making and safety.

Revolutionizing AI Deployment With Baseten On Hugging Face Inference Providers

Hugging Face now supports Baseten for conversational and text-generation workloads, offering users more deployment options with no immediate performance data released.

The Ultimate Guide To Grok Bot For Engineering Applications

xAI reveals a new Grok-based assistant tailored for engineering tasks, but details on features, availability, and performance remain undisclosed.

The cleaner cap table. Why Anthropic’s public-benefit structure dodges OpenAI’s charitable-trust problem — and trades it for a governance question of its own.

Analysis of how Anthropic’s mission-focused, trust-based governance differs from OpenAI’s conversion history, impacting their IPO prospects.