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

🔍 Read the full analysis: Could AI's Self-Improvement Become Recursive? Reading The Evidence 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

Recent assessments indicate that AI is accelerating parts of AI research and development, with humans still setting goals and supervising work. Anthropic and independent evaluator METR say fully autonomous development of a successor model has not been demonstrated; proposed U.S. legislation to restrict the process remains in committee.

A viral essay by podcaster Dwarkesh Patel has renewed debate over whether AI is beginning to improve itself, but the evidence cited so far points to AI-assisted research with humans in control, not systems independently designing their successors. Anthropic and evaluator METR report faster progress on some AI-development tasks, while a proposed U.S. ban on recursive self-improvement remains stalled in committee.

In a November 13 essay, Patel argued that AI is already speeding up the work of building AI and forecast that progress could accelerate further. The essay draws on assessments and figures published by AI companies and METR; its predictions are Patel’s analysis, not new experimental results. The distinction matters because “recursive self-improvement” can describe anything from AI tools helping researchers to a system independently designing and developing its replacement, as explored in Anthropic’s evidence on recursive self-improvement.

Anthropic’s August 2026 figures say AI is involved in more than 90% of its research and development work at a level described as collaboration or above, and that Claude “leads” 26% of the work. The company says no measured subset operates fully autonomously. These figures are Anthropic’s own assessment, using an automation scale that distinguishes collaboration from AI leading work under human supervision.

METR, a nonprofit evaluator, examined internal Anthropic data from August 12 to October 29. In a November 5 report, it said the company was likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025. METR attached high uncertainty to that estimate, saying the actual acceleration might be closer to 1.3 or 3 times. It attributed most of the likely speed-up to AI assistance, but said AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. METR also said Anthropic redacted much of the evidence it considered most informative.

Separate figures attributed to OpenAI offer another view, though they have not received independent review according to The Decoder, which reported on the company’s September blog post. OpenAI said it had met its measured goal of an “automated research intern” and aims for a fully automated AI researcher by March 2028. The reported results show a gap between shorter and longer tasks: tasks taking under 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.

At a glance
analysisWhen: Developing; assessments and legislative…
The developmentA viral essay has renewed debate over recursive AI self-improvement, while current assessments describe AI-assisted research rather than autonomous 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

How AI Changes Research Work

AI assistance could shorten parts of the research cycle, allowing teams to test ideas, write code, or analyze results more quickly. If that assistance keeps improving, it may affect how fast new models are developed and how much research can be done with a given workforce. That is a meaningful change even without a system that can independently create its successor.

The distinction also shapes public debate and policy. Treating AI-assisted work as proof of autonomous self-improvement overstates what the cited evidence establishes; treating the current evidence as insignificant would overlook reported gains in research productivity. For readers, the practical question is not only whether a feedback loop exists, but how much of the process AI performs, how reliably it does so, and who retains authority over goals, training and deployment.

The proposed legislation could reach beyond a narrow definition of systems that design their own successors. Its text would pause training, modification and fine-tuning of large AI systems until a new Department of Artificial Intelligence sets rules. It would also cover systems capable of automating or greatly accelerating AI research and development. That breadth makes the definitions and any eventual regulatory framework consequential for companies and research institutions.

Products Worth Considering

Jackery Explorer 1500 v2 Portable Power Station with 200W Solar Panel

Jackery Explorer 1500 v2 Portable Power Station with 200W Solar Panel

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

As an affiliate, we earn on qualifying purchases.

The Bills and Their Status

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 the following day. The proposals would pause specified work on large AI systems, including recursive self-improvement, pending rules from a new federal department. As of the status records described in the source material, both bills had been referred to committee and had not advanced; the House version had 12 Democratic cosponsors.

A second proposal, Rep. Ro Khanna’s “Human Control Over AI Act,” was reported by CNBC in September as a summary. The source material says no bill text could be found, so its scope cannot be assessed on the same basis as the Sanders-Casar bills. The renewed attention follows Democrats’ gains in both chambers in the November 3 elections, which supporters of a ban view as an opportunity in the next Congress. That political interpretation is not itself evidence that legislation will pass.

Anthropic describes the endpoint of recursive self-improvement as an AI system that can fully and autonomously design and develop its own successor. Its formulation helps separate that possibility from current work in which people choose objectives and supervise AI contributions. The available examples describe activity within companies, not a publicly documented case of a model independently building a successor.

““We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.””

— Anthropic, in its essay “When AI builds itself”

Limits of the Current Evidence

The available assessments do not establish that any AI system can independently design and develop its own successor. Anthropic’s figures are company-reported, while METR’s evaluation used internal data and included substantial redactions. METR itself emphasized high uncertainty around the estimated pace of improvement.

OpenAI’s reported results also come from the company, with no independent review identified by The Decoder. The figures concern different tasks and measures, so they do not provide a direct comparison with Anthropic’s assessment. It remains unclear how much of the reported acceleration comes from AI tools rather than other changes in staffing, computing resources, research methods or project selection.

The legal picture is also unsettled: the Sanders-Casar bills have not advanced beyond committee referral, and the available material on Khanna’s proposal is only a summary. The source does not establish whether either proposal will be reintroduced, amended or taken up in the next Congress.

Next Tests and Congressional Steps

Further independent evaluations could clarify whether AI’s contribution to model development is growing, which tasks it can complete without intervention, and whether gains persist across companies and research settings. Future assessments will be more informative if they provide transparent methods and enough evidence for outside researchers to check the results.

OpenAI’s stated target is a full automated AI researcher by March 2028, following its claimed automated-intern milestone. That is a company goal, not a confirmed outcome. In Congress, the next developments would be committee action or new bill text; until then, the proposals do not change requirements for businesses or users. The central factual question remains whether AI can move from assisting human-led research to reliably directing and completing the creation of its own successors.

Key Questions

Has recursive AI self-improvement been demonstrated?

Not in the fully autonomous sense described by Anthropic: a system independently designing and developing its own successor. Current reports describe AI assisting researchers, with humans retaining oversight.

What did METR find about Anthropic’s progress?

METR estimated that Anthropic was improving model capabilities on well-specified tasks at around twice its earlier rate. It marked the estimate as highly uncertain and said AI assistance likely explained most of the acceleration.

Would the proposed U.S. ban affect AI companies now?

No. The Sanders-Casar bills had been referred to committee and had not advanced, according to the status records cited in the source material. The proposals currently impose no new requirements.

Why is the evidence disputed or incomplete?

Much of the data comes from the companies being assessed. METR relied on internal Anthropic data and said many informative materials were redacted; The Decoder reported that OpenAI’s figures had no independent review.

Source: ThorstenMeyerAI.com

You May Also Like

Microsoft’s Staff Letter And Buyout Offer: A Closer Look At AI At Work

Microsoft offered voluntary buyouts to about 60,000 employees. A staff letter and reported worker accounts raise questions about AI pressure and morale.

How We Check Frontier Models For Signs Of Misalignment

OpenAI has a page titled “How we detect frontier model misalignment,” but its methods, findings and publication date could not be verified.

Steam App 1905180 Climbing The Steam Charts

The Steam application identified as 1905180 has climbed to rank 17 on Steam’s most-played charts, reaching a peak of 30,926 players. The cause remains unconfirmed.

Using Watch-Once Commands To Automate Desktop Tasks Efficiently

New approach enables users to record desktop workflows once and replay via spoken commands, boosting automation efficiency for power users.