🔍 Read the full analysis: Has AI Begun Improving Itself? The Evidence And Policy Questions on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
AI systems are already assisting with AI research and development, and Anthropic and independent evaluator METR report evidence that the assistance may be speeding progress on some tasks, as explored in Anthropic’s evidence on recursive self-improvement. The available evidence does not show an AI independently designing and building its successor; a proposed U.S. bill to restrict such development has not advanced beyond committee referral.
AI is helping researchers develop AI systems, but the evidence available as of November 16 does not show a system independently designing and building its own successor. The question returned to public debate after podcaster Dwarkesh Patel published an essay on November 13; company disclosures and an independent assessment describe faster AI-assisted work, while a proposed congressional ban remains in committee.
Anthropic reported that, as of August 2026, Claude led 26% of the company’s AI research and development work, meaning it completed most of some tasks from a high-level prompt while a person supervised. The company said more than 90% of the measured work involved AI at least collaborating, but that no measured subset was fully autonomous. Those figures describe Anthropic’s own work and measurements, not an industry-wide rate.
The nonprofit evaluator METR assessed Anthropic using internal data from August 12 to October 29. In a report dated November 5, it said Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025. METR stressed that estimate had high uncertainty: it said the actual acceleration could be closer to 1.3 or 3 times. It judged that AI assistance likely accounted for most of the increase, while saying AI had not yet made highly significant, fully autonomous contributions to the most important parts of model training.
METR also said Anthropic had redacted much of the evidence it considered most informative. OpenAI separately reported reaching an internal milestone it called an “automated research intern,” according to The Decoder’s account of OpenAI’s September blog post. The outlet reported that OpenAI said short tasks succeeded 86% of the time without intervention, while more than half of successful tasks lasting four to eight hours of human work required at least one human step. These are company-reported results; The Decoder said it found no independent review cited by OpenAI.
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
Anthropic: “We are not there yet.”
- 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.
- AI leads about a quarterIn 26%, Claude completes most of a task from a high-level prompt while a human supervises.
- No closed loopNo measured part runs fully autonomously. That is the step “recursive self-improvement” usually means.
- 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
AI Assistance, Not Independent Research
The distinction matters because AI-assisted research can speed development without amounting to autonomous self-improvement. If models help write code, run experiments or analyze results, human researchers may complete more work or improve systems faster. But people can still choose research goals, decide which results to pursue and control training and deployment. That is different from an AI system independently repeating the full cycle to produce a successor.
The evidence also comes with limits. Anthropic’s figures describe its own work, and METR’s assessment relies on internal company data with substantial redactions. OpenAI’s figures, as reported by The Decoder, likewise rely on the company’s measurements. The reports offer indications of acceleration, not a settled measurement of industry-wide progress or proof that humans have been removed from the process.
The policy stakes are immediate even without a fully autonomous system. Lawmakers are debating restrictions that could cover not only systems that build successors, but also tools that substantially accelerate AI research. Depending on how those terms are defined, legislation could reach present-day research practices as well as more advanced systems. The current evidence makes the wording and scope of any such rules consequential for labs, workers and users.
Popular right now

DREAME D20 Air Plus Robot Vacuum and Mop

Hisense 43-Inch E6 Cinema Series Hi-QLED 4K UHD Smart Fire TV (43E6QF)

KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Juniper (KSM150PS)
As an affiliate, we earn on qualifying purchases.
The Bills Behind the Debate
Anthropic describes recursive self-improvement as an AI system autonomously designing and developing its own successor. Its essay, “When AI builds itself,” says “We are not there yet” and adds that the outcome is not inevitable, though it could arrive sooner than institutions are prepared for. Patel’s November 13 essay, “The avalanche that hit math is coming for AI development,” draws on METR’s assessment and figures published by Anthropic and OpenAI; its forecasts are Patel’s analysis, not new experimental findings.
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 on September 24. The bills would pause training, modification and fine-tuning of large AI systems until a new Department of Artificial Intelligence set rules. Their text includes systems with “the capacity to automate or greatly accelerate” AI research and development, as well as recursive self-improvement. Both bills have been referred to committee and have not advanced; the House version has 12 Democratic cosponsors, according to the source material.
A separate proposal, Rep. Ro Khanna’s Human Control Over AI Act, was described in a September CNBC report as a summary. The source material says no bill text was available. The issue has gained political attention following the November 3 elections, when the Associated Press called both chambers of Congress for Democrats. No change to business or user requirements follows from the introduced bills while they remain unpassed.
““We are not there yet, and recursive self-improvement is not inevitable.””
— Anthropic, in its essay “When AI builds itself”
How Much Work Is Truly Automated?
It remains unclear how much of the reported acceleration is caused by AI assistance, and how well the findings apply beyond the companies studied. METR characterized its estimate as highly uncertain and noted that much of its most informative evidence was redacted. The public material does not provide enough detail to independently reproduce the assessment.
There is also no single agreed threshold for when AI-assisted research becomes “recursive self-improvement.” A model that helps with selected coding or analysis tasks is not the same as one that sets goals, directs experiments and builds a successor without human intervention. The available reports describe activity in the middle of that spectrum. No evidence in the source material establishes a fully autonomous successor-building loop, and forecasts about when one might appear remain forecasts.
The proposed legislation’s reach is also unsettled. The bills include language about systems that greatly accelerate AI research, but the source material does not establish how regulators would interpret that phrase or which current tools could fall under it. Neither bill has become law, and no operational rules have been set.
Committee Action and Better Evidence
The immediate legislative milestone is whether the House or Senate committees receiving the bills hold hearings, revise the text or advance it. As of November 16, both measures remain at the committee-referral stage; the source material identifies no scheduled vote. The separate Khanna proposal would require bill text before its scope and status could be assessed on the same basis.
On the research side, further public data and independent evaluation would help clarify whether AI assistance is producing sustained acceleration, which tasks benefit, and how much human supervision remains necessary. OpenAI has stated a goal of a full automated AI researcher by March 2028, according to The Decoder’s account, but that is a company target, not a confirmed milestone. For now, the evidence supports a narrower conclusion: AI is contributing to AI research, while fully autonomous development of successor models has not been documented in the material available.
Key Questions
Has AI started improving itself?
AI is assisting people who develop AI, and METR and company reports indicate that this may be speeding some research tasks. The available evidence does not show an AI independently designing and building its own successor.
What did METR find about Anthropic?
METR said Anthropic was likely improving model capabilities on well-specified tasks at around twice its earlier rate, while stressing high uncertainty. It judged that AI assistance probably explained most of the increase and said AI had not made highly significant, fully autonomous contributions to the most important parts of model training.
What would the Ban Artificial Superintelligence Act do?
The proposed bills would pause training, modification and fine-tuning of large AI systems until a new federal department set rules. They also cover systems capable of automating or greatly accelerating AI research and development. They have not passed and remain referred to committee.
When does OpenAI expect a full automated AI researcher?
According to The Decoder’s account of OpenAI’s September blog post, the company has set a goal of reaching a full automated AI researcher by March 2028. That date is a company target, not an independently verified forecast or completed milestone.
Source: ThorstenMeyerAI.com
