Recursive Self-Improvement In AI: 3 Questions About The Evidence And A Ban
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Anthropic says Claude leads 26% of its AI research and development work, while no measured portion runs fully autonomously. A proposed federal bill would pause some advanced AI development, including systems able to greatly accelerate AI research, but both versions remain in committee and the bill does not define how to measure that threshold.

Anthropic reports that its Claude system leads 26% of the company’s AI research and development work, but says no measured portion of that work is fully autonomous—a distinction at the center of renewed debate over recursive self-improvement. The debate resurfaced after podcaster Dwarkesh Patel published an essay on November 13 arguing that AI is already speeding up AI development, while a proposed federal ban on some advanced AI work remains in Congress without committee action.

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.

Anthropic’s figures, reported for August 2026, use an automation scale from AL0, no AI involvement, to AL5, fully autonomous work with no human in the loop. The company says more than 90% of its AI R&D work is at or above the “AI collaborates” level. It classifies Claude as leading 26% of that work, meaning it completes most of a task from a high-level prompt while a person supervises. Anthropic says it has measured no portion of its AI R&D work at the fully autonomous level.

The independent evaluator METR judged that Anthropic’s progress on well-specified tasks had accelerated to about twice its earlier rate, attributing that mainly to AI assistance rather than fully autonomous contributions. Patel’s essay draws on that assessment and figures published by Anthropic and OpenAI; the essay’s forecast of much faster progress is Patel’s interpretation, not a new empirical finding.

The Ban Artificial Superintelligence Act of 2026 was introduced by Sen. Bernie Sanders as S.5493 on September 23 and Rep. Greg Casar as H.R.10538 the following day. It would pause training, modification and fine-tuning of large AI systems until a proposed Department of Artificial Intelligence sets rules. The text also targets systems with “the capacity to automate or greatly accelerate” AI research and development. As of November 16, both bills are referred to committee; the House bill has 12 Democratic cosponsors.

At a glance
reportWhen: Essay published November 13, 2026; bill…
The developmentA viral essay by Dwarkesh Patel renewed debate about whether AI is improving the process of building AI, as a congressional bill proposing limits on that activity remains stalled in committee.
Graphic titled Who says it, and what it rests on. Four cards. METR, highlighted, outside check: around twice the earlier pace at Anthropic, range 1.3x to 3x, primarily via AI assistance, most key evidence redacted. Anthropic, own data: Claude leads 26 percent of its AI R&D work in August 2026, no measured part fully autonomous. OpenAI, own data: automated research intern reached according to our measurements, as The Decoder reports, no independent review named. Dwarkesh Patel, commentary: essay of November 13, built on the three above; his timeline is his own forecast, not a measurement
What the claim that recursive self-improvement has started rests on. Sources: METR, Nov 5, 2026; Anthropic; The Decoder, Sep 7, 2026, on OpenAI's post; dwarkesh.com, Nov 13, 2026. Graphic: Thorsten Meyer AI.

Where the Bill Draws Its Line

The dispute affects how policymakers might regulate AI systems that help develop future systems. AI assistance and autonomous self-improvement are different capabilities: the reported evidence describes AI doing substantial work under human supervision, while the bill’s language reaches systems able to automate or greatly accelerate AI R&D. That distinction matters for labs, workers and users because a broad interpretation could cover tools used in ordinary research, while a narrower one might focus on systems that can carry out development with little human direction.

The legislation does not specify a measurement for “greatly accelerate.” The available Anthropic figures describe work at one company, and the source material identifies no standard method for deciding when AI has crossed that statutory threshold. The policy question is therefore not settled by the 26% figure alone. It turns on how the terms are defined, what evidence regulators use and whether the proposed agency would set limits on research tools or only more autonomous systems.

Products Worth Considering

Bose QuietComfort Headphones - Wireless Bluetooth Over-Ear Noise Cancelling Headphones, Petal Pink Limited Edition

Bose QuietComfort Headphones – Wireless Bluetooth Over-Ear Noise Cancelling Headphones, Petal Pink Limited Edition

As an affiliate, we earn on qualifying purchases.

From Lab Metrics to Congress

Recursive self-improvement generally refers to a feedback loop in which an AI system helps create a more capable AI system, which then helps build its successor. Anthropic defines the endpoint as an AI system able to fully autonomously design and develop its own successor. The company has said that endpoint has not been reached and that such progress is not inevitable, though it could arrive sooner than institutions are prepared for.

Patel’s November 13 essay and narrated episode renewed attention to the question, but the evidence it cites predates the essay and includes company-reported figures. The legislative push also predates the essay: Sanders introduced S.5493 on September 23, and Casar introduced H.R.10538 on September 24. The source material links supporters’ renewed hopes for action to Democrats winning both chambers in the November 3 election, but the bills’ recorded status remains referral to committee.

Rep. Ro Khanna has also proposed a “Human Control Over AI Act,” but the available source describes it only as a summary CNBC saw in September. It reports that no bill text was located, so its specific scope and legal effect cannot be assessed from the material available.

““Claude is not operating fully autonomously for any measured subset of AI R&D work.””

— Anthropic

The Threshold Is Still Undefined

The available evidence does not establish that an AI system has independently designed and built its own successor. Anthropic’s reported measurements cover its own AI R&D work as of August 2026; the source material does not provide comparable measurements across other labs. METR’s assessment concerns progress on well-specified tasks and attributes the acceleration primarily to AI assistance.

It is also unclear how the proposed legislation would distinguish routine AI-supported research from work that “greatly accelerate[s]” AI development. The bill text, as described in the source, does not define that term or set a measurement standard. Khanna’s proposal remains less assessable because the source reports a summary but no bill text. Neither bill has advanced beyond committee referral in the status information provided.

Committee Action Would Set the Next Test

The next formal milestone for S.5493 and H.R.10538 would be committee consideration, where lawmakers could debate definitions, scope and enforcement. No hearing, vote or timetable is identified in the source material. If either bill advances, its treatment of AI systems that speed up research—and the standard used to measure that effect—will determine how its proposed pause applies in practice.

For now, the evidence and the law remain at different stages: researchers and companies are reporting degrees of AI assistance, while Congress has proposed but not enacted a restriction. Further measurements across labs, clearer definitions and legislative action would be needed to establish whether the feared autonomous loop has emerged or how the proposed ban would apply.

Key Questions

Has AI begun recursively improving itself?

The reported evidence shows AI assisting and leading parts of research under human supervision. Anthropic says no measured portion of its AI R&D work is fully autonomous, and the source material does not document an AI independently building its successor.

What does Anthropic say Claude does in its AI research?

For August 2026, Anthropic reports that more than 90% of its AI R&D work is at or above the “AI collaborates” level, and that Claude leads 26%. The company says no measured subset operates fully autonomously.

What would the Ban Artificial Superintelligence Act do?

The introduced bills would pause training, modification and fine-tuning of large AI systems until a proposed Department of Artificial Intelligence sets rules. They also address systems able to automate or greatly accelerate AI research and development.

Has the proposed ban become law?

No. As of November 16, 2026, S.5493 and H.R.10538 were referred to committee. The source material reports no committee action or enactment.

Source: Thorsten Meyer AI

You May Also Like

AI Regulation Tightens: The Significance Of The August 2 Deadline Change

The EU AI Act’s high-risk enforcement date was delayed to December 2027, but transparency obligations remain effective from August 2, 2026. Here’s what you need to know.

The Channel Move: Anthropic, Wall Street, and the Acquisition of the Real Economy

Anthropic, Blackstone, Goldman Sachs, and others launch a $1.5 billion joint venture to embed AI into thousands of portfolio companies, transforming enterprise deployment.

Can A 512GB Mac Studio Support Frontier AI? What ‘Run’ Really Means

Exploring whether Apple’s new Mac Studio with 512GB memory can run frontier-scale AI locally, and what ‘run’ truly entails for performance and capacity.

How SAP’s €1 Billion AI Spend Is Transforming Data Tables Instead Of Chatbots

SAP has completed a €1 billion acquisition of Prior Labs, emphasizing enterprise data table AI, signaling a shift from chatbots to structured data models.