Recursive Self-Improvement In AI: Evidence, Definitions, And Policy
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Recursive Self-Improvement In AI: Evidence, Definitions, And Policy on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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

A November essay by Dwarkesh Patel renewed debate over whether AI is already accelerating AI research, drawing on company disclosures and a METR assessment. The documented work involves human oversight; autonomous systems designing their own successors have not been shown. Two proposed U.S. bills would restrict advanced AI development, but neither has advanced beyond committee referral.

Debate over recursive self-improvement in AI intensified after podcaster Dwarkesh Patel published an essay on November 13 arguing that AI is already helping speed up AI development. The evidence cited in the essay describes AI-assisted research under human direction, not a system autonomously building its own successor; two proposed congressional bills that would restrict AI development remain at the committee stage.

Recursive self-improvement, or RSI, generally describes a feedback loop in which an AI system helps create a more capable system, which then contributes to further development. The central disagreement is about where to draw the line between AI assisting researchers and AI independently carrying out that loop. Anthropic defines the more advanced endpoint as an AI system that can fully autonomously design and develop its own successor, and says that has not been reached.

Anthropic reported that in August 2026 Claude led 26% of the company’s AI research and development work. More than 90% of the work was at or above a level where AI collaborates, but Anthropic said no measured subset was fully autonomous. These are company-reported figures, not an independent demonstration of an AI system designing and training a successor without human involvement.

In a November 5 assessment based on internal data from August 12 to October 29, the nonprofit METR estimated that Anthropic was improving model capabilities on well-specified tasks at about twice its rate from March 2024 to October 2025. METR attached high uncertainty, saying the actual acceleration could be closer to 1.3 times or three times. It judged that AI assistance likely accounted for most of the speed-up, while saying AI could not yet make 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.

Patel’s essay draws on METR’s assessment and figures published by AI companies; its forecasts are his own interpretation, not new experimental findings. Separately, The Decoder reported on OpenAI’s September blog post after its website blocked the article’s download. According to The Decoder, OpenAI said it had reached an “automated research intern” milestone by its own measurements, while aiming for a fully automated AI researcher by March 2028. The report said short tasks often succeeded without intervention, but more than half of successful tasks lasting four to eight human hours needed at least one human step. OpenAI said people still set priorities and make decisions about scaling, pauses and deployment.

At a glance
reportWhen: Bills introduced September 23-24, 2026;…
The developmentA viral essay renewed attention to evidence of AI-assisted research and two congressional bills that would restrict development of advanced AI systems.
Crypto market snapshot
Fear & Greed Index
10/100 — Extreme Fear
Bitcoin BTC$64,883▲ 0.7%
Ethereum ETH$1,818▲ 1.3%
Tether USDT$0.9995▲ 0.0%
BNB BNB$593.85▲ 1.1%
USDC USDC$0.9997▲ 0.0%
XRP XRP$0.9654▲ 1.8%
Solana SOL$71.9▲ 1.7%
TRON TRX$0.3037▼ 1.1%
Live data · CoinGecko · alternative.me (24h change)

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

Why AI Research Speed Matters

If AI tools help researchers produce more capable models faster, development timelines and the demands on safety evaluation could change even before systems can act independently. The METR estimate suggests a possible acceleration at one company, but its stated uncertainty and reliance on internal, partly redacted evidence limit what can be concluded about the wider industry.

The debate also has immediate policy relevance. The Ban Artificial Superintelligence Act of 2026 would pause specified AI development activities until a new Department of Artificial Intelligence established rules. Its scope includes systems capable of automating or greatly accelerating AI research and development. The proposal could therefore cover more than a narrow definition of an autonomous system that designs its own successor. Neither bill has become law, and the proposals do not change requirements for businesses or users today.

Products Worth Considering

LG UltraGear 32GS60QC-B 32-Inch QHD 180Hz Curved Gaming Monitor

LG UltraGear 32GS60QC-B 32-Inch QHD 180Hz Curved Gaming Monitor

Apple MacBook Pro 14.2-inch with Tandem OLED Display, Space Black

Apple MacBook Pro 14.2-inch with Tandem OLED Display, Space Black

As an affiliate, we earn on qualifying purchases.

Two Bills, Limited Progress

Sen. Bernie Sanders introduced the Ban Artificial Superintelligence Act as S.5493 on September 23; Rep. Greg Casar introduced its House counterpart, H.R.10538, on September 24. The bills would pause the training, modification and fine-tuning of large AI systems, including through recursive self-improvement, until the proposed department set rules. Their latest recorded status is referral to committee. The House measure had 12 cosponsors, all Democrats.

A separate 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 was available to review, so its precise scope cannot be compared with Sanders and Casar’s legislation. The proposals returned to discussion after Democrats won both chambers in the November 3 election, according to the source account; supporters view the next Congress as a possible opening for action.

Anthropic’s automation scale helps explain why descriptions of RSI differ. It runs from no AI involvement to fully autonomous operation. At the “leads” level, AI completes most of a task from a high-level prompt while a human supervises; at the highest level, the system operates without a human in the loop. Evidence of AI leading some research tasks does not by itself establish the latter.

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

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

Limits of the Current Evidence

The available material does not establish that an AI system has independently designed, built and deployed its own successor. Company figures describe internal work and use different measures; METR’s estimate carries high uncertainty and relies on evidence that the evaluator said was substantially redacted. OpenAI’s reported measurements were also company-generated, according to The Decoder’s account.

It is also unclear how much the reported productivity gains would generalize beyond the tasks and organizations studied, or how the proposed legislation would define and enforce its restrictions if enacted. The House and Senate bills remain in committee, while the available information about Khanna’s proposal is only a summary. Patel’s predictions about faster future progress remain forecasts, not established outcomes.

Committee Action and New Evidence

The next legislative indicator is whether either committee takes up S.5493 or H.R.10538; the source material reports no scheduled vote or further action. Khanna’s proposal would need published bill text for its provisions to be assessed directly. Until then, none of these measures changes current rules.

On the technical side, further independent evaluations and clearer disclosure of methods and evidence would help test whether AI is accelerating research, how broadly that effect applies, and whether systems are moving from supervised assistance toward autonomous work. For now, the strongest supported conclusion is narrower: AI is contributing to parts of AI development, while researchers retain control over goals and key decisions.

Key Questions

Has recursive self-improvement begun?

That depends on the definition. AI is assisting with and leading some research tasks under human supervision, but the cited evidence does not show an AI system independently designing and developing its own successor.

What did METR find about Anthropic’s research pace?

METR estimated that Anthropic’s progress on well-specified tasks was about twice its rate from March 2024 to October 2025. It called the estimate highly uncertain, giving a possible range closer to 1.3 times or three times, and said AI assistance likely explained most of the speed-up.

What would the proposed federal ban do?

The Ban Artificial Superintelligence Act would pause training, modification and fine-tuning of large AI systems, including through recursive self-improvement, until a proposed Department of Artificial Intelligence set rules. The bills also cover systems capable of automating or greatly accelerating AI research and development.

Have the bills passed?

No. The Senate and House versions were referred to committee and have not become law. The source material reports no scheduled vote.

Does the evidence show AI researchers are fully autonomous?

No. Anthropic said no measured subset of its AI research and development work was fully autonomous. METR also said AI could not yet make highly significant, fully autonomous contributions to the most important parts of model training.

Source: ThorstenMeyerAI.com

Nothing in this article is financial or investment advice. Cryptocurrency and precious-metal investments carry significant risk — do your own research and consider a licensed advisor.
You May Also Like

Claude’s Game Recompilation Skills May Have Been Quietly Adjusted

A Kotaku headline suggests Claude’s game reverse-engineering ability may have declined, but the available material offers no tests or Anthropic response.

Maximize Mobile Productivity With These AI-Enhanced Laptops In 2026

Discover the top AI-optimized laptops in 2026 designed to boost mobile productivity, featuring advanced hardware and intelligent features for professionals.

When One Agent Isn’t Enough: Claude Now Builds Its Own Team of Agents on the Fly

Anthropic’s Claude now creates its own team of agents dynamically, improving handling of complex, high-value tasks through new workflow orchestration.

The Report On Zhou Chang’s Startup Plans And ByteDance’s Response

ByteDance denied a report that Seed AI research leader Zhou Chang left to start a company; his employment status and plans remain unverified.