AIThis post was created with the assistance of artificial intelligence (AI).

By Thorsten Meyer | November 16, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Has recursive self-improvement started? This week the question went viral. Podcaster Dwarkesh Patel published an essay, which he also narrates as an episode, arguing that AI is already speeding up the making of AI, and that much faster progress may follow. At the same time, a bill sits in Congress that would ban that loop, with penalties of up to 20 years in prison. Both sides use the same words. They do not always mean the same thing. So we checked two questions at the source: what has actually been shown, and what the proposed ban would actually cover.

A disclosure up front: Anthropic and OpenAI are central to this story, and Anthropic makes Claude, which helped research and draft this article.

Key takeaways

  • “Started” depends on the definition. The documented part is AI that assists with research and leads parts of it while humans choose the goals. Anthropic reports that as of August 2026 Claude “leads” 26% of its AI research and development work, and that no measured part runs fully autonomously. METR, an independent evaluator, judged that Anthropic’s progress on well-specified tasks runs at around twice the earlier rate, “primarily via AI assistance rather than fully autonomous contributions.” A model that designs and builds its own successor on its own is not documented. Anthropic: “We are not there yet.”
  • The viral claim is commentary, not new evidence. Patel’s essay (November 13) rests on METR’s assessment and on figures that Anthropic and OpenAI published themselves. The rest is his own forecast, which he labels as such.
  • The ban is real, and it has bill text. The Ban Artificial Superintelligence Act of 2026 was introduced by Sen. Bernie Sanders (S.5493, September 23) and Rep. Greg Casar (H.R.10538, September 24). It would pause the training, modification and fine-tuning of large AI systems, “including through recursive self-improvement,” until a new Department of Artificial Intelligence sets rules. It would also ban systems with “the capacity to automate or greatly accelerate the process of artificial intelligence research and development.”
  • It has not moved. The latest status records show both bills referred to committee, with 12 House cosponsors, all Democrats. A second proposal, Rep. Ro Khanna’s “Human Control Over AI Act,” exists only as a summary that CNBC saw in September. We found no bill text.
  • Why it is back in the news: the Associated Press called both chambers of Congress for the Democrats on November 3, and ban supporters see a new chance in the next Congress. Nothing changes for businesses or users today.

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

Products Worth Considering

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

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

Amazon Kindle Paperwhite Signature Edition (2024), 32 GB, Metallic Black

Amazon Kindle Paperwhite Signature Edition (2024), 32 GB, Metallic Black

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

As an affiliate, we earn on qualifying purchases.

What people mean when they say it has started

Recursive self-improvement, often shortened to RSI, describes a loop: an AI system helps make a better AI system, which then helps make the next one. Anthropic, in its essay “When AI builds itself,” defines the end point as “an AI system capable of fully autonomously designing and developing its own successor.” It adds: “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”

Between humans doing everything and AI doing everything lies a wide middle. Anthropic measures it with an automation scale from Epoch AI that runs from AL0, no AI involvement, to AL5, where AI “operates fully autonomously, with no human in the loop.” In AL3, AI “collaborates.” In AL4, AI “leads”: it completes most of a task from a high-level prompt while a human supervises. Anthropic’s figures for August 2026:

  • The share of work at or above “AI collaborates” is above 90%.
  • Claude “leads” 26% of Anthropic’s AI research and development work.
  • “Claude is not operating fully autonomously for any measured subset of AI R&D work.”

So the honest answer to the question whether it has started is: the loop has started to turn with humans in it. The version that worries people most, where the human steps out, has not been shown anywhere we could find.

What the evidence actually is

METR, November 5. The nonprofit evaluated Anthropic from the inside, with internal data, from August 12 to October 29, 2026. Its central finding: “As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.” It gives that with “high uncertainty” and “would not be surprised if the true acceleration was closer to 1.3x or 3x.” It thinks a majority of the speed-up is likely due to Anthropic’s use of AI. It also says “AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process,” while researchers described “signs of life.” And it says plainly that “Anthropic redacted a majority of the evidence we found most informative for these conclusions.” We covered the report in detail when it came out.

OpenAI, September. OpenAI published its own internal figures in a blog post. Its website blocked our download, so we rely on The Decoder, which read the post. According to The Decoder, OpenAI says it reached its goal of an “automated research intern,” with the milestone met “according to our measurements,” and it wants a full automated AI researcher by March 2028. The median researcher uses more than $600 a day of inference at API prices; the 90th percentile, above $7,000. Tasks under 15 minutes succeeded 86 percent of the time without any intervention, but for successful tasks of four to eight human hours, more than half needed at least one human step. OpenAI says people still set research priorities and decide on scaling, pauses and deployment. The Decoder notes that all the data comes from inside the company and that OpenAI mentions no independent review.

Dwarkesh Patel, November 13. The essay behind the viral video is titled “The avalanche that hit math is coming for AI development.” Patel cites the same METR finding and the same company figures. His own view, clearly marked as a forecast: “I’ve now come to the view that we’re only a couple of years away, and possibly only a couple of months away, from AIs that are as good as the best humans at AI research.” He sketches two outcomes, a “software-only singularity” and a more limited “spark,” and gives each a 50% chance. He also lists the strongest objections, among them that automated researchers would still be “bottlenecked by compute for experiments,” and writes: “these bottlenecks are serious.” That is one well-read commentator’s forecast. It is not a measurement, and we do not treat it as one.

Google DeepMind. Its CEO Demis Hassabis talked about self-improving AI in a recent podcast interview. The transcript page and a report about it blocked our downloads, so we do not quote him here.

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.

What the ban would cover

The Ban Artificial Superintelligence Act is a 19-page bill, according to NBC News, and its text is on the government’s own bill site. It has four parts that matter for this debate.

1. A pause for large systems. An “advanced artificial intelligence system” is one “trained on computing power greater than or equal to” 10^25 integer or floating point operations, a threshold the new Secretary would adjust each year. Each such system would be under a “mandatory pause”: it “may not be trained, modified, or fine-tuned, including through recursive self-improvement,” and no unreleased system may be deployed. The pause ends only when the Secretary determines that the new Department “is fully staffed” and has issued its rules.

2. A ban on superintelligence and its “precursors.” The bill defines artificial superintelligence as a system that exceeds human performance “across most domains or tasks,” or that could plan and carry out “the destruction or disempowerment of humanity.” It also bans systems with “superintelligence precursor characteristics.” Two of the six are about self-improvement: “(A) The capacity to automate or greatly accelerate the process of artificial intelligence research and development” and “(E) The capacity to independently modify or enhance its own functions.” An aide involved in drafting told NBC News that the goal is to bar recursive self-improvement.

3. Shutdown. A system identified with precursor characteristics would be paused and “sequestered from the internet.” If the Secretary cannot verify within 30 days that the characteristics are gone, it is “rendered inoperative.” Anyone who discovers such a system must report it within 24 hours.

4. Licences and penalties. Developing or distributing an advanced model would need a charter from the Secretary. Policymaking staff who recklessly break the rules could face up to 20 years in prison. Other employees could be barred from the AI industry for 10 years, and a company that loses its charter could forfeit its intellectual property and assets.

Graphic titled What the Ban Artificial Superintelligence Act would do, bill S.5493 and H.R.10538. Pause large systems trained with at least 10 to the power of 25 operations: no training, modifying or fine-tuning, including through recursive self-improvement, until a new department sets rules. Highlighted: ban precursors, including the capacity to automate or greatly accelerate AI R&D and to independently modify or enhance its own functions. Shut down: cut off from the internet, rendered inoperative after 30 days unless cleared, report finds within 24 hours. Licence and punish: a charter to develop or distribute, up to 20 years in prison for policymaking staff. Status: in committee, 12 House cosponsors
The Sanders-Casar bill in four parts, from the bill text. Source: S.5493 and H.R.10538 via govinfo; bill status record of Oct 14, 2026. Summary, not legal advice. Graphic: Thorsten Meyer AI.

Where it stands. The Senate bill was referred to the Committee on Commerce, Science, and Transportation, the House bill to the Committee on Oversight and Government Reform. The House version had 9 original cosponsors and 12 by October 13, all Democrats, according to the official status record, last updated October 14. No hearing or vote is recorded. After the November 3 elections, Rep. Yassamin Ansari, one of the cosponsors, called on X for a pause and a permanent ban on superintelligence, as Gadget Review reported. The new Congress starts in January.

The second proposal. On September 28, CNBC reported that Rep. Ro Khanna “will introduce” a “Human Control Over AI Act.” According to the summary CNBC saw, it “would ban models that recursively self-improve or autonomously modify their own core objectives, containment or shutdown controls until federal guardrails exist and an agency approves the activities.” It is a conditional ban, not a permanent one, and Khanna said it is modeled on conversations with safety groups including METR. We searched the official bill databases and found no bill with that name, no number and no text. Until text appears, its details are reported, not confirmed.

And the court route. Separately, four nonprofits asked a California court on November 12 to stop five AI developers from building more capable systems and to forbid AI conducting or directing research into better models. We covered the lawsuit on November 14. No court has ruled.

Graphic titled Two bills and a lawsuit, at very different stages. Highlighted: the Sanders and Casar bill of September 23 and 24, a permanent ban on superintelligence and its precursors and a pause for large systems, text published, in committee. The Khanna proposal of September 28, a ban on self-improving models until an agency approves, per a summary CNBC saw, no bill text found. The nonprofits' lawsuit of November 12, asking a California court to cap capability and bar AI-directed research into better models, no ruling
Three routes to stopping recursive self-improvement, and where each stands. Sources: govinfo; CNBC, Sep 28, 2026; Public Citizen, Nov 12, 2026; Decrypt. Graphic: Thorsten Meyer AI.

The arguments for and against

For a ban. Sen. Sanders: “When you are racing towards a cliff, you don’t just ease up on the gas pedal. You hit the brakes.” Rep. Casar said the bill would halt “the capacity for AI to develop new AI instead of humans.” The bill’s findings quote warnings from industry leaders, among them Anthropic’s CEO Dario Amodei and OpenAI’s CEO Sam Altman. The advocacy group ControlAI praised the choice “to define superintelligent AI by its capability to overpower the federal government,” as NBC News reported.

Against, or not like this. Caleb Max of the National AI Association, a business-friendly trade group, told NBC News: “Some CEOs would say we have already crossed the threshold set out in the bill, meaning this is a de facto AI ban”. He called the penalties “likely unconstitutional and, again, poorly defined.” Kevin Frazier of the University of Texas School of Law welcomed the push to act but said the bill ignores other paths, such as independent verification organizations.

What the labs themselves propose. We found no statement by Anthropic, OpenAI or Google DeepMind on the bill itself. The bill’s findings claim that Anthropic, OpenAI and xAI “agree that humanity should slow down development of artificial intelligence” and “pace the frontier”; we did not check that claim at the companies’ own sources. Their published proposals are about measurement and oversight, not a ban. Anthropic suggests that frontier developers publish automation measures regularly, verified by third parties, which “could also become the trigger for stronger requirements, like a fixed testing window before a new model is used for further AI R&D.” OpenAI’s chief scientist Jakub Pachocki wrote, according to The Decoder: “Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” He calls for binding standards and international coordination, while OpenAI keeps working toward an automated researcher.

Our reading: the hard part is the line

This is our reading of the bill text, not legal advice.

The bill’s broadest trigger is not a self-building AI. It is “the capacity to automate or greatly accelerate” AI research. That wording does not say how much is “greatly,” or who measures it. METR’s “around twice,” measured with company data that is mostly redacted, is exactly the kind of evidence such a test would need, and exactly the kind that is hard to check from outside. The trade group’s “already crossed” argument and the bill’s own definitions point at the same gap: the law would turn on a measurement that nobody yet makes in a standard way.

The pause is also broader than the RSI debate. As written, it would cover any training, modification or fine-tuning of a system above the compute threshold, not just loops in which AI improves AI.

What it means for businesses

  • Today: nothing changes. The bills sit in committee; there is no law and no court order. Every model and API keeps working.
  • Plan for faster cycles either way. If AI is speeding up AI research, new model versions come faster. Budget time for re-testing your own use cases, not just for licences.
  • If you fine-tune large models, watch the definitions. In the Sanders-Casar text, the pause applies to modifying and fine-tuning covered systems, and a charter would be needed to develop or distribute them. Whether a future version keeps that scope matters more for most companies than the RSI clause.
  • Ask vendors for evidence, not adjectives. METR’s report shows the honest form: a headline number, a range, the parts that are redacted. OpenAI’s milestone is reported “according to our measurements.” Ask who checked.

What it means for users

  • Nothing you use stops working. No bill has passed, and the lawsuit has no ruling.
  • Read claims that it has started carefully. It is true for AI that helps researchers. It is not shown for AI that improves itself alone.
  • The debate is about control as much as speed. Both the bill’s supporters and the labs’ own safety leads say oversight has not kept up. They disagree on whether the answer is a ban, a licence or better measurement.

What is not known

  • The redacted evidence. METR says most of what it found most informative was redacted. Nobody outside Anthropic and METR knows what it shows.
  • The Khanna bill. We found no text, number or introduction record.
  • The next Congress. Whether the new majority takes up either proposal, and in what form, is open.
  • How “greatly accelerate” would be measured. The bill leaves it to a department that does not exist yet.
  • OpenAI’s own post and the Hassabis interview, which we could not download, are cited here only through other reports or not at all.

The bottom line

The claim that recursive self-improvement has started is true in a narrow, documented sense: AI already does a large share of the work of building AI, under human direction, and at least one lab’s pace has roughly doubled. The full loop, where an AI builds its successor on its own, is not documented. A real bill would ban both the full loop and, arguably, the narrow version, but it is sitting in committee, and its key line depends on a measurement nobody makes in a standard way yet. The useful next step is not a louder headline. It is the measurements that would tell us which side of that line we are on.

I’m Thorsten Meyer. The full sources are below, and more reality checks are at thorstenmeyerai.com.

Related on this site: When AI Starts Improving AI, what METR found inside Anthropic, Terminal-RSI, checked and the nonprofits’ lawsuit.

Sources

  • US Congress: S.5493, Ban Artificial Superintelligence Act of 2026, introduced in the Senate September 23, 2026, text via govinfo. govinfo.gov
  • US Congress: H.R.10538, Ban Artificial Superintelligence Act of 2026, introduced in the House September 24, 2026, text and bill status record (updated October 14, 2026). govinfo.gov · congress.gov
  • Office of Rep. Greg Casar: press release on the Ban Artificial Superintelligence Act, September 2026. casar.house.gov
  • NBC News, Sahil Kapur and Jared Perlo: Bernie Sanders and Greg Casar propose AI ‘superintelligence’ ban with a 20-year jail penalty, September 23, 2026. nbcnews.com
  • Semafor: Progressive Democrats push superintelligence ban, September 23, 2026. semafor.com
  • CNBC: Khanna to introduce AI safety bill with ban on ‘recursive’ technology until safeguards exist, September 28, 2026. cnbc.com
  • Spectrum News, citing AP race calls: Democrats regain control of the Senate and House, November 2026. spectrumlocalnews.com
  • Gadget Review via Yahoo News: Congresswoman Calls for AI Superintelligence Ban as Democrats Prepare to Take Control of Congress, November 5, 2026. yahoo.com
  • Public Citizen: Coalition of Nonprofits Files Landmark Lawsuit to Stop Development of Advanced AI Systems at Anthropic, Google, Meta, OpenAI, and SpaceXSI…, press release, November 12, 2026, and Decrypt’s report on the complaint. citizen.org · decrypt.co
  • METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, November 5, 2026. metr.org
  • Anthropic (The Anthropic Institute): When AI builds itself and Measurements for understanding the pace of AI development inside frontier labs, 2026 (copies saved on November 7, 2026). When AI builds itself · Measurements
  • The Decoder, Maximilian Schreiner: OpenAI reports AI ‘research interns’ and warns about its own pace at the same time, September 7, 2026 (on OpenAI’s post Research acceleration: The view inside OpenAI, which blocked our download). the-decoder.com
  • Dwarkesh Patel: The avalanche that hit math is coming for AI development, November 13, 2026. dwarkesh.com
  • Not accessed (download blocked; not quoted): the Masters of Scale interview with Demis Hassabis and Inc.’s report on it; OpenAI’s post itself; the Sanders section-by-section summary PDF.

Researched and drafted with AI assistance, including Anthropic’s Claude. Anthropic and OpenAI are among the companies this article discusses; neither they, METR, nor any lawmaker named here reviewed or sponsored it. We report each side’s position as written, stated or reported and take no view on whether the bill should pass. This article is for information only and is not legal or investment advice. Sources checked on November 16, 2026. Our own text and graphics contain no affiliate links; offer and product boxes added by the site may contain Amazon affiliate links. Images: AI-generated illustration (Runware), animation and graphics by Thorsten Meyer AI; no real people or logos are shown in our images. Boxes added by the site may show product photos or people.

You May Also Like

Anthropic Says You Can’t Be Mean To Claude. It’s Not About Feelings – CNN

Anthropic says hostile prompts may affect Claude’s behavior and safety. The available report does not establish that the chatbot has feelings.

Tesla’s New Roadster Is a Farewell to the Human Driver. That’s the Real Announcement.

Tesla’s second-generation Roadster flips, jumps and fires SpaceX thrusters. The bigger news is how Tesla described it: “the best of the last human-driven cars.” Its capital now goes to Cybercab, FSD, Optimus and AI compute.

Reality Check: Will Every Job Really Be Automated?

Facing the automation wave, many wonder which jobs will truly disappear and which will adapt—discover the surprising realities ahead.

We Gave 42 AI Setups the Same 15 Legal Questions. The Model Built for Law Scored 8 Out of 100

On Lexassay, my legal AI test, 42 AI setups answered the same 15 questions on German, US and EU law without internet. Z.ai’s GLM-5.3 FlashX Legal, built for law, scored 8 of 100; Europe’s Mistral Large 4 scored 10; Claude Opus 5.5 led with 79. Another test ranks them almost the other way round. What a wrong legal answer can cost, and how to protect yourself. Disclosure: Thorsten Meyer AI runs Lexassay; Claude agents grade. Not legal advice.