How To Assess Claims That AI Has Started Improving Itself
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Anthropic says Claude led 26% of its AI research and development work in August 2026, while no measured part ran fully autonomously. An essay by Dwarkesh Patel forecast faster progress, but the proposed federal bills that cite recursive self-improvement remain in committee.

AI is helping researchers build AI, but available evidence does not show systems independently designing and developing their own successors. Anthropic reported that Claude led 26% of its AI research and development work in August 2026, with a human supervising and no measured portion operating fully autonomously; the figures have renewed attention after a viral essay and proposed US legislation targeting AI self-improvement.

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.

In a November 13 essay and podcast episode, Dwarkesh Patel argued that AI is already accelerating the development of AI and could lead to much faster progress. The essay draws on assessments by the independent evaluator METR and figures published by Anthropic and OpenAI. Patel presents forecasts about what could follow as his own analysis; the essay does not report a new demonstration of autonomous self-improvement.

Anthropic’s August 2026 measurements describe several levels of AI involvement in its research work. The company says more than 90% of that work involves AI collaboration or a higher level of involvement, and Claude leads 26% of the work by completing most of a task from a high-level prompt while a person supervises. Anthropic says it measured no subset of AI research and development operating fully autonomously.

METR assessed that Anthropic’s progress on well-specified tasks was running at about twice its earlier rate, primarily through AI assistance rather than fully autonomous contributions. That finding concerns the pace of progress on a defined class of tasks; it does not establish that AI systems are independently building successor systems.

At a glance
reportWhen: Essay published November 13, 2026; bill…
The developmentA viral essay renewed debate over whether AI is improving the process of building AI, as two proposed US bills seek to restrict systems that automate or greatly accelerate AI research.
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 AI Assistance Becomes a Policy Question

The distinction matters because claims of self-improving AI can refer to different things: AI helping researchers with tasks, AI leading work under human supervision, or an autonomous system designing its own successor. The evidence cited here supports the first two at Anthropic, while the fully autonomous loop remains undocumented.

That boundary also matters for policy. The Ban Artificial Superintelligence Act of 2026 would restrict training, modification and fine-tuning of large AI systems and target systems with the capacity to automate or greatly accelerate AI research and development. Because the bill uses a broad acceleration standard, deciding which research activity falls within it could affect both labs and the pace of AI development. The bills have not become law, and the source material says nothing changes for businesses or users today.

Products Worth Considering

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Ice Blue

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Ice Blue

As an affiliate, we earn on qualifying purchases.

From Research Metrics to Proposed Limits

Anthropic describes recursive self-improvement as an AI system fully autonomously designing and developing its own successor. Its essay, “When AI builds itself,” says that point has not been reached and that such progress is not inevitable, though it could arrive sooner than institutions expect. The company’s measurements use an automation scale from AL0, with no AI involvement, to AL5, with no human in the loop. The reported 26% figure corresponds to AI leading work under human supervision, not to AL5 autonomy.

Sen. Bernie Sanders introduced S.5493 on September 23, 2026, and Rep. Greg Casar introduced H.R.10538 the following day. The source material reports that both bills were referred to committee and had not advanced; the House measure had 12 Democratic cosponsors. It also describes a separate proposal from Rep. Ro Khanna, the Human Control Over AI Act, based on a summary reported by CNBC in September. The article’s source says it found no bill text for that proposal.

The debate returned to wider attention after the Associated Press called both chambers of Congress for Democrats in the November 3 elections. Supporters of the proposed restrictions see the next Congress as a possible opening to advance them, but that political outlook is not evidence that the bills will pass.

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

— Anthropic

The Acceleration Threshold Is Undefined

The bills describe systems able to “greatly accelerate” AI research and development, but the source material says the legislation does not define how much acceleration counts as “greatly.” It also says no standard measurement currently settles where that threshold lies. That leaves uncertainty over how the proposed restriction might apply to AI tools that assist or lead research under human supervision.

The reported Anthropic metrics come from the company itself and cover its own AI research and development work as of August 2026. The material does not establish whether other labs see the same levels of AI involvement. Nor does it document an AI system completing the fully autonomous successor-design loop. Patel’s forecast of faster progress remains an argument about what may happen, rather than confirmation that it has happened.

Committee Decisions and Better Measurements

The next formal step for S.5493 and H.R.10538 would be action in their committees. The source material reports no committee movement or scheduled votes. Any consideration of the proposals would need to address how regulators could measure “greatly accelerate” and distinguish supervised AI assistance from autonomous research activity.

For the underlying technology question, more transparent and comparable measures across labs could clarify how much work AI performs, how closely people supervise it, and whether any system can independently develop a successor. Until such evidence appears, the supported conclusion is limited: AI is contributing to research that builds AI, while fully autonomous recursive self-improvement has not been demonstrated in the cited material.

Key Questions

Has AI started improving itself?

AI is helping researchers develop AI, and Anthropic says Claude led 26% of its AI research and development work in August 2026 under human supervision. The cited material does not show a system autonomously designing and developing its own successor.

What did Anthropic report about Claude?

Anthropic reported that more than 90% of its AI research and development work involved AI collaboration or a higher level of involvement. It said Claude led 26% of the work, but no measured subset ran fully autonomously.

What did METR conclude?

METR assessed that Anthropic’s progress on well-specified tasks was running at around twice its earlier rate, primarily through AI assistance rather than fully autonomous contributions.

Would the proposed US bills ban AI research?

The bills would pause training, modification and fine-tuning of large AI systems until a proposed Department of Artificial Intelligence set rules. They also target systems with the capacity to automate or greatly accelerate AI research and development. Both remain in committee, and the source material says the acceleration threshold is undefined.

Source: Thorsten Meyer AI

You May Also Like

1.08 Million Email Addresses Leaked In Osaka—Here’s What To Do

Osaka says 1,078,713 user email addresses were exposed through an email vendor. The city says messages and application data were not leaked.

What Signals Reveal Misalignment In Frontier AI Models?

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

Microsoft’s Patch Tuesday Update Tackles 1,094 Flaws And An Attack

Microsoft’s November 10 update lists 1,094 CVEs, including an Outlook flaw reported as exploited before its fix. A separate Paint issue remains.

Robotics And AI Factories Get An Open Library Of Simulation-Ready 3D Assets

SimReady Central offers more than 2,000 OpenUSD assets for robotics simulation and industrial digital twins; licensing and performance details remain unclear.