TL;DR
Anthropic says Claude leads 26% of its AI research and development work, while no measured part runs fully autonomously. A pair of congressional bills would restrict AI training and systems that greatly accelerate AI research, but both remain in committee and key terms are undefined.
Anthropic reports that AI leads 26% of its research and development work, but says none of the measured work is fully autonomous—a distinction at the center of renewed debate over whether recursive self-improvement has begun. The debate intensified after podcaster Dwarkesh Patel published an essay on November 13 arguing AI is already speeding up AI development, while two bills in Congress propose restrictions on systems that automate or greatly accelerate AI research.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic’s measurements, covering its AI R&D as of August 2026, use an automation scale developed by Epoch AI. The company says more than 90% of the work falls at or above “AI collaborates”, where AI handles substantial tasks under close human direction. It says Claude “leads” 26% of the work, meaning the model completes most of a task from a high-level prompt while a person supervises. Anthropic reports no measured subset operates fully autonomously.
The independent evaluator METR assessed that Anthropic’s progress on well-specified tasks was running at around twice its earlier rate, “primarily via AI assistance rather than fully autonomous contributions.” Patel’s essay draws on that assessment and figures published by Anthropic and OpenAI. His forecast of much faster progress is commentary, not a new measurement. Anthropic defines the endpoint of recursive self-improvement as an AI system that can fully autonomously design and develop its own successor; it says, “We are not there yet.”
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 proposed Department of Artificial Intelligence establishes rules. They also propose a ban on systems with “the capacity to automate or greatly accelerate” AI research and development. Both bills have been referred to committee. The House version lists 12 Democratic cosponsors.

Where AI Oversight Could Begin
The distinction between AI assisting researchers and AI autonomously building a successor matters because the two capabilities pose different questions for oversight. Anthropic’s figures indicate substantial AI involvement in its own R&D, but the company’s measurements do not document a fully autonomous development loop. METR’s assessment likewise attributes faster progress primarily to AI assistance, not autonomous contributions.
The proposed bills could reach systems before that fully autonomous threshold. Their language includes the capacity to “greatly accelerate” AI research, so a system could potentially fall within their scope even if people still set goals and supervise work. That makes the definition and measurement of acceleration consequential for labs, users and policymakers. The bills have not advanced, and nothing in the source material indicates a change to what businesses or users may do today.
Products Worth Considering

LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor

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

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From Assistance to Autonomy
“Recursive self-improvement” is used for a loop in which an AI system helps develop a more capable AI system, which in turn contributes to further development. The term can describe a range of human involvement. Anthropic’s Epoch AI scale runs from AL0, with no AI involvement, to AL5, where AI operates fully autonomously with no human in the loop. Its reported “collaborates” and “leads” categories describe intermediate levels with human direction or supervision.
Patel’s November 13 essay and narrated episode brought these claims to wider attention. The proposed restrictions predate the essay: Sanders introduced the Senate bill on September 23, and Casar introduced the House version the following day. A separate proposal from Rep. Ro Khanna, called the Human Control Over AI Act, was reported by CNBC in September as a summary; the source material says no bill text was found.
The source article links renewed interest in a potential ban to the November 3 election, after which the Associated Press called both chambers for Democrats. Supporters see an opening in the next Congress, but that political outlook is not evidence that the bills will pass or that their provisions will be adopted.
““Claude is not operating fully autonomously for any measured subset of AI R&D work.””
— Anthropic
The Bill’s Acceleration Test
“Greatly accelerate” is not defined in the bill language described by the source, and no standard measurement method is identified. It is therefore unclear which systems would meet that threshold, including systems that still depend on human oversight. Anthropic’s figures are the company’s own measurements of its work; they do not establish the level of automation across other labs.
It also remains unclear whether either bill will receive further consideration. Both are in committee, and the source provides no later action or timetable. Khanna’s proposal has been described only through a summary, without bill text available to assess its scope.
Committee Action and Measurement
The next legislative milestone would be action on S.5493 or H.R.10538 in their respective committees; no hearing, vote or schedule is specified in the source material. Any future debate will depend in part on how lawmakers define and measure a system’s ability to accelerate AI research.
For the technical question, the evidence to watch is whether labs or independent evaluators document AI systems taking on more of the research process without human direction. The current figures describe substantial AI assistance and some AI-led work at Anthropic, while leaving a fully autonomous successor-design loop unobserved in its measurements.
Key Questions
Has recursive self-improvement been demonstrated?
Not in the fully autonomous sense described by Anthropic. The company reports AI-led work with human supervision, but no measured subset of its AI R&D operates fully autonomously.
What does Anthropic say Claude does in its research?
As of August 2026, Anthropic says Claude “leads” 26% of its AI research and development work. The company also says more than 90% of that work is at or above the “AI collaborates” level; these are Anthropic’s measurements.
What would the proposed federal bills restrict?
The Senate and House versions would pause training, modification and fine-tuning of large AI systems until a proposed federal department sets rules. They also target systems with the capacity to automate or greatly accelerate AI research and development.
Have the bills become law?
No. The source reports that S.5493 and H.R.10538 were referred to committee and had not advanced further at the time described.
What is still uncertain about the bills?
The bill language does not define “greatly accelerate,” and the source identifies no standard way to measure it. The bills’ eventual prospects and the scope of a separate proposal from Rep. Ro Khanna also remain unclear.
Source: Thorsten Meyer AI
