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

🔍 Read the full analysis: Is Recursive AI Improvement Underway? Evidence And Possible Limits on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • 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

Evidence indicates that AI tools are accelerating some AI research and development, with humans still setting goals and supervising work. Anthropic and an independent evaluator say the process is not fully autonomous; the viral claim that recursive self-improvement is underway depends on how the term is defined. Bills that would restrict such systems remain in committee.

AI is helping speed some AI research, but available evidence does not show a system independently building its own successor, according to Anthropic and independent evaluator METR. The distinction is at the center of renewed debate after podcaster Dwarkesh Patel published an essay on November 13 arguing that AI is already accelerating the development of AI, while federal bills proposing limits on the process remain stalled in Congress.

Anthropic reported that in August 2026, Claude led 26% of the company’s AI research and development work, using a scale where “leading” means completing most of a task from a high-level prompt while a human supervises. The company said more than 90% of its measured work involved AI at the level of collaboration or higher. It also said Claude was not fully autonomous in any measured subset of its AI research and development.

METR, a nonprofit evaluator, assessed Anthropic’s internal data from August 12 to October 29. In a report published November 5, it estimated that Anthropic was improving model capabilities on well-specified tasks at around twice the rate seen from March 2024 to October 2025. METR attached high uncertainty to that estimate, saying the true acceleration could be closer to 1.3 times or three times. It judged that AI assistance was likely responsible for most of the speed-up, while saying AI had not yet made highly significant, fully autonomous contributions to the most important parts of model training.

Patel’s essay draws on METR’s assessment and figures released by Anthropic and OpenAI. His account of what may happen next is forecast and interpretation, not new experimental evidence. OpenAI has separately described reaching its goal of an “automated research intern,” according to a post summarized by The Decoder. The outlet reported that short tasks often succeeded without intervention, while more than half of successful tasks lasting four to eight human hours required at least one human step.

At a glance
reportWhen: Developing; key assessments and comment…
The developmentA viral essay has renewed debate over recursive AI improvement as company disclosures and an independent evaluation document AI assistance in research, but not fully autonomous development.

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

Human Oversight Still Shapes AI Research

The evidence points to a meaningful change in how AI companies conduct research: models can help with tasks that feed into more capable models, and an independent evaluator found a faster pace on some well-defined tasks at Anthropic. If that assistance continues to improve, it could affect the speed and cost of AI development, as well as how much work human researchers need to do.

That is different from the stronger claim that a model can autonomously design, train and deploy its replacement. Human researchers still select priorities, supervise work and make decisions about scaling, pauses and deployment, according to the company descriptions cited in the source material. The distinction matters for public debate and policy: AI-assisted research is already reported; a self-directed improvement loop remains unproven.

The question also has immediate policy relevance. Bills introduced in September would restrict the training or modification of certain large AI systems, including systems capable of accelerating AI research. Yet those proposals have not advanced out of committee, so they do not currently change what businesses or users can do.

Shark Pet Cordless Stick Vacuum, 40-Minute Runtime, Lightweight, IX141

Shark Pet Cordless Stick Vacuum, 40-Minute Runtime, Lightweight, IX141

As an affiliate, we earn on qualifying purchases.

How the Debate Reached Congress

Recursive self-improvement, or RSI, generally describes a feedback loop in which an AI system helps create a more capable AI system, which then assists with the next one. Anthropic defines the stronger endpoint as a system that can fully autonomously design and develop its own successor. The company has said that endpoint has not been reached and is not inevitable.

The current evidence sits between ordinary software assistance and that autonomous endpoint. Anthropic uses a scale from Epoch AI that ranges from no AI involvement to full autonomy. On that scale, “collaborates” means a model works alongside a person; “leads” means it handles most of a task with human supervision. Anthropic’s August figures describe substantial AI involvement, but not the removal of people from the research process.

METR’s findings are based on an evaluation using Anthropic’s internal data, rather than a public, independent replication of every underlying result. The evaluator said Anthropic redacted much of the evidence it found most informative. Separately, Patel’s November 13 essay brought these developments to a wider audience. Its viral reach has renewed attention to evidence that was already published, rather than documenting a newly demonstrated autonomous system.

The legislative debate is also active but unresolved. The Ban Artificial Superintelligence Act of 2026 was introduced by Sen. Bernie Sanders as S.5493 on September 23 and by Rep. Greg Casar as H.R.10538 on September 24. The proposed legislation would pause certain training, modification and fine-tuning until a new Department of Artificial Intelligence sets rules. As of the latest status cited, both bills were referred to committee; the House bill had 12 Democratic cosponsors.

““We are not there yet, and recursive self-improvement is not inevitable.””

— Anthropic, in its essay “When AI Builds Itself”

Limits in the Available Evidence

The central uncertainty is partly definitional: “recursive self-improvement” can describe a range of human-AI workflows, from AI assistance to a system independently making its successor. The sources support the first kind, but do not document the second. Calling the entire process “underway” without explaining which meaning is intended risks blurring that difference.

METR’s estimate is not a precise measure of how much AI caused the faster research pace. The evaluator described its conclusion as highly uncertain and noted that much of the evidence it considered most informative was redacted. The findings also concern well-specified tasks; they do not establish that AI can direct the full research and development process.

OpenAI’s figures, as reported by The Decoder, are company-provided measurements. The source material says no independent review was cited, and some longer tasks still involved human intervention. The available information does not establish whether similar results hold across other companies, research settings or less clearly defined work. It is also unclear how quickly today’s assistance might advance toward more autonomous contributions.

The legislative outlook is uncertain as well. The two Sanders and Casar bills remain at committee referral, and the source reports no further movement. A separate proposal associated with Rep. Ro Khanna was described in a September CNBC report, but the source says it could not find bill text. Neither proposal should be treated as enacted law.

Research and Bills Face Further Tests

The next useful evidence would be independent evaluations that can be reproduced, clearer disclosures about which research tasks AI completes, and results showing how much human intervention is needed as tasks become longer or less structured. New assessments would also help establish whether the acceleration METR estimated persists beyond the period it examined.

For now, Anthropic’s reported measurements and METR’s evaluation describe AI supporting research under human direction. OpenAI’s reported target of a full automated AI researcher by March 2028 is a stated goal, not proof that the milestone will be reached. The source material does not provide a confirmed date for a new independent assessment or a congressional vote.

On Capitol Hill, the next milestone would be committee action on the Sanders and Casar bills, or the introduction of formal text for any separate proposal. Until such action occurs, the proposals remain pending. The broader question—whether AI assistance will develop into a self-sustaining loop—will depend on evidence of systems taking on more of the research process, not on predictions alone.

Key Questions

Has recursive AI self-improvement started?

AI assistance in AI research is documented, including Anthropic’s report that Claude led 26% of its measured AI research and development work in August 2026. A system fully autonomously designing and developing its own successor has not been documented in the cited evidence.

What did METR find about Anthropic’s research pace?

METR estimated that Anthropic improved model capabilities on well-specified tasks at around twice its earlier rate, comparing August 2026 with March 2024 to October 2025. It called the estimate highly uncertain and said the true acceleration could be nearer 1.3 times or three times.

Are the proposed US restrictions in force?

No. The Ban Artificial Superintelligence Act of 2026 was introduced in the House and Senate, but the latest status in the source material shows both bills referred to committee. The proposals have not become law.

What is the main evidence gap?

The available reports do not show AI independently directing the full development of a successor system. METR also reported substantial redactions in the evidence it reviewed, while OpenAI’s figures were described as internal measurements without cited independent review.

Source: ThorstenMeyerAI.com

You May Also Like

Does US Law Let Employees Opt Out Of Workplace AI?

A University of Kentucky employee received an accommodation from personally using AI. The decision was individual, not a court ruling or broad opt-out right.

Are VPNs Legal Technical Tools? EU Court Says Yes In Landmark Ruling

The EU Court has confirmed that VPNs are lawful technical tools, marking a landmark ruling that clarifies their legal status in the EU.

Decoding The $400 Million AI Public Option: Sovereign Infrastructure Or Political Show?

An analysis of the $400 million public-interest AI initiative launched by France and global partners, examining its progress, funding, and implications.

The Safety Card, Played From Every Side: David Sacks, Anthropic, and the Fable Standoff

White House official claims Anthropic refused to fix a cyberweapon jailbreak, leading to model bans; Anthropic disputes this, citing minor flaws. Details remain unclear.