COBUS KOK

Tracker · companion to Brakes Without a Moat · Draft 3 · Beta

Brake or Moat?

The AI rules and proposals on the table now, graded with one test, each with a reason and a date

Every AI rule on the table gets called safety by one side and capture by the other. You can say that about any rule, so it settles nothing. In Brakes Without a Moat I proposed a test that reads what a rule does, whoever asked for it. This page runs it on the rules and proposals people are arguing about now. Each row says what the rule stops and gives a verdict, with one reason. The last column is the date I graded it.

The rules, graded

Where the essay already graded a rule, I use its verdict word for word and mark it (essay). Some proposals have no text yet. Others have text that leaves the deciding answer to an agreement still to come. I grade those Too early to tell, or Undecided where the essay graded them, and the note says what would decide the grade. The test's default is to assume a moat. That applies once there is text and I still can't tell. Nothing here is aimed at concentration so far. The nearest candidate is the EU's move to put Amazon's and Microsoft's clouds under its rules for gatekeepers. I'll grade it when the decision is final. On 11 November 2026 I rechecked four rows against their sources: the three steps of the pacing proposal and New York's RAISE Act. The other rows stand as graded on the dates shown.

RuleWhoWhat it stopsVerdictGraded
Pacing, step one: embedded evaluatorsAnthropic; OpenAI pledged the sameModels shipping on the lab's word aloneA brake in design (essay). Since October an evaluator that one lab doesn't pay has worked inside it. It has published three reports, all from work that began earlier. The lab still decides what is left out, and for its newest models the outside testers' findings came as the lab's own summary.11 Nov 2026
Pacing, step two: a common paceDario Amodei; welcomed in general terms by Sam Altman, Elon Musk and Demis HassabisThe speed of the frontier, by agreement among the leading labsUndecided (essay), until it says where its line sits and who checks it. It becomes a brake only if the line binds frontier runs alone and someone the labs don't pay checks it.11 Nov 2026
Pacing, step three, and the US–China incident lineAmodei; the US–China incident channel announced on 25 SepAI help with bioweapons first; each government warned of the other's serious incidentsThe right direction (essay). The only part that does not depend on the lead. The incident channel is a first piece, if it grows to cover capability jumps. It is still being set up.11 Nov 2026
Antitrust safe harbour for labs (S. 5105)Senators Schiff and Banks; stalled with the defence billNothing itself: it lets labs agree to delay or limit AIToo early to tell. It names a checker the labs don't pay, the Justice Department. It names no line, so the verdict belongs to whatever the labs agree.23 Sep 2026
A standards body on the FINRA modelDemis Hassabis of Google DeepMind; OpenAI, Anthropic and Google in talksFrontier releases that fail its tests, voluntary at firstToo early to tell. It would bind only frontier models, which the test allows. But the labs it checks would pay for it.23 Sep 2026
A federal liability shield for labsSought by labs, as reported; the Treasury Secretary said noNothing a model could do: it moves the cost of harmA trade. It spends the public's right to sue, so I oppose it on its own. At most it could be the price of a real brake, one with a checker the labs don't pay.27 Sep 2026
Ban Artificial Superintelligence ActSenator Sanders and Representative Casar, introduced 23 SepTraining and new releases at 10^25 operations and up, and foreign models without approval, until a new department writes its rulesA brake, but only inside the US (essay). It stops a capability, spares the small and has a public checker. But it fails the fourth question. It stops qualifying runs at home and none abroad, and keeps foreign models out with a wall.11 Oct 2026
Senate frontier AI draftSenators Thune, Cruz and Klobuchar; no text yetReleases the government judges unsafe, as reportedToo early to tell. As reported, the labs test themselves and the Commerce Secretary approves. So it turns on who checks.23 Sep 2026
FRONTIER Act (H.R. 9925)Representatives Obernolte, Trahan and four others, introduced 23 JulFrontier releases without a report; unaudited practice at the biggest labsBrake. It has a compute line for everyone, and the heavy duties fall only on firms spending $1 billion or more on AI. Publishing weights counts as release. Its text would also preempt the core of SB 53 and New York's RAISE Act, both graded brakes here.23 Sep 2026
California SB 53California; in force since 1 Jan 2026Frontier releases without a published framework and report; unreported incidentsA brake, but nobody outside audits the labs yet (essay). A startup training a small model faces nothing, and the labs report on themselves to the state.11 Oct 2026
New York's RAISE Act, as amendedNew York; registration announced 10 Nov, duties from Jan 2027The same as SB 53, with incidents reported within 72 hoursBrake. SB 53's two lines, and a state office funded by the largest developers but not run by them.11 Nov 2026
EU AI Act: models with systemic riskThe EU; the AI Office has been able to enforce it since 2 Aug 2026Models above 10^25 operations without evaluation, adversarial testing and incident reportsBrake. The Commission must keep its compute line current, and the checker is one the labs don't pay. Open weights are bound the same as closed ones.23 Sep 2026
Entity-based regulationDean Ball and Ketan Ramakrishnan, July 2025Unsafe practice at firms spending over $1 billion a year on AI researchBrake. A rule on entities is a brake when they are defined by what they spend, and this one is.23 Sep 2026
US export controls on AI chipsThe US Commerce DepartmentChinese buyers getting the best American chipsA wall. It stops a country rather than a capability or a company. This year's rules have chased the routes around it.23 Sep 2026
Bans on open-weight modelsDiscussed in Washington since July; nothing writtenAnyone downloading and running the weightsA trade of less misuse for more power in fewer hands, at too high a price (essay). I'd refuse it, but it should be argued as a trade.11 Oct 2026

Notes and sources

Pacing, step one. On 12 September Dario Amodei proposed outside evaluators inside the lab. They'd have desks, badges and access close to the lab's own risk team's, and be free to publish key findings. The lab may redact security, legal and commercial material, but not findings because they are unfavourable. Sam Altman pledged the same for OpenAI that day. On 18 September Anthropic named its first evaluator, Accenture. Anthropic said it will pay for the work directly for now, with pooled or government money as the aim. A lab paying its evaluators makes them independent the way a paid auditor is. That helps, but it's weaker than a checker the lab doesn't pay. On 15 October Anthropic said six people from METR now have company hardware, office space and access to internal tools "roughly on par with what our own risk assessment teams have". It said it "can't veto publication", but can require METR to remove information in categories set in advance. METR says of the contract, "No fees are paid in either direction." Anthropic described the Accenture work that day as something it was still working toward. METR has since published reviews of two sections of Anthropic's August risk report, on 20 October and 5 November, and an assessment of how fast the lab is improving its models, on 5 November. All three came from work that began before the October arrangement was announced. The 20 October review was written under earlier terms: its authors did not have that access, Anthropic could have vetoed it, and METR says nothing was redacted. The two November reports came out of eleven weeks in which Anthropic shared confidential material with METR, and Anthropic could have material removed from both. In the assessment METR says Anthropic "redacted a majority of the evidence we found most informative". The nearest thing so far to an evaluator's own report on a model is METR's summary of its testing of Claude Opus 5.5, published on 22 September, the day Anthropic released that model. METR was not paid for it, and Anthropic could review and edit the text. On 9 November Anthropic released new models and said outside evaluators, among them the UK AI Security Institute and METR, had tested them first. Their findings appeared as a summary in Anthropic's system card, which the evaluators did not review. At OpenAI, the company said on 9 October it was still finalising contracts with its outside assessors. On 22 October METR and Redwood Research published a short account of three weeks spent testing OpenAI's agent monitoring in July and August, under terms that let OpenAI remove anything it judged sensitive. Regraded on 11 November, and the verdict stays. On the test's third question the answer is better at one lab, where the checker is now one the lab doesn't pay. It is still a brake in design because the lab decides what is redacted, and at the 9 November release the evaluators' findings reached the public through the lab. Sources: Amodei; Anthropic, 18 Sep; Anthropic, 15 Oct; METR on Opus 5.5, 22 Sep; METR, 20 Oct; METR on OpenAI, 22 Oct; METR's review and assessment, 5 Nov; Anthropic, 9 Nov; TechCrunch, 12 Sep; CBS News, 9 Oct.

Pacing, step two. The leading labs would agree common safety standards and limits on the rate of unchecked progress. The government would grant "a narrow waiver" so they can talk without breaking competition law. An agreement among the largest competitors to slow down together raises an antitrust question. It becomes a brake only when it says where its line sits and who checks it. If its limits bind only frontier runs, a startup training a small model faces nothing, which the test allows. If signing becomes the price of deploying at all, it is a licence. Standards checked by evaluators the labs pay are a closed loop. The waiver's likeliest vehicle stalled with the defence bill over the summer. Senators Hawley and Cruz have said no to any antitrust exemption for AI (see the safe harbour). And on 18 September four subscribers sued four labs under the Sherman Act. They argue that an agreement to slow down restricts output, which puts the antitrust question to a court. All four labs' lawyers had entered appearances by 12 October. The labs' answers were due on 14 and 15 October, and on 12 October OpenAI filed a stipulation to extend its time to respond. The public copy of the docket, last refreshed that day, shows no answer and no ruling on the claims. SpaceXAI has since asked the court to send the claims against it to arbitration, and the plaintiffs have opposed that, MLex reported on 28 October. The first case management conference is set for 23 December. Rechecked on 11 November: the proposal still names neither a line nor a checker, so the verdict stays. Sources: Amodei; Bloomberg Law on Buist v. Anthropic; the docket on PacerMonitor; MLex, 28 Oct.

Pacing, step three, and the incident line. Amodei's third step is coordination with China. It starts with a ban on AI help for biological weapons. If trust allows, it moves on to testing before release and a limit on the speed of self-improvement. On 20 September, after talks with China's vice premier, the US Treasury Secretary said the two sides had discussed a formal AI dialogue. He also said the US had proposed a way for the two governments to warn each other of AI incidents serious enough to touch national security. On 25 September the White House said the two countries had set up the dialogue and agreed "to establish a bilateral communication channel for SI incidents", SI being the administration's new name for AI. It said the next exchange would occur by November 2026. China's Foreign Ministry has backed the dialogue in public and has not mentioned an incident channel. Neither side has published what counts as an incident. In remarks the Treasury published on 11 November, the Treasury Secretary said "we are establishing a direct channel" for serious incidents, and that the two sides "expect to exchange best practices for evaluating and testing advanced models before their release". He also said the two presidents would meet in Shenzhen the following week. By 11 November I had found no announcement from either government that the next exchange has taken place. As announced, the channel covers incidents, not capability. It becomes the first disclosure brake between rivals if it grows to cover capability jumps. Rechecked on 11 November, and the verdict stays. Sources: Amodei; Euronews, 21 Sep; White House, 25 Sep; US Treasury, 11 Nov.

Antitrust safe harbour. The Collaboration on Adversarial Threats and Security Risks Act was introduced on 23 July. It would exempt two things from antitrust law. One is sharing information about AI security risks. The other is agreements "delaying or otherwise limiting" the release, training or testing of AI, after written notice to the Justice Department's antitrust chief naming the risk. The sharing exemption excludes price-fixing, market allocation and boycotts. Firms claiming either exemption must prove good faith and a security purpose, and the Attorney General can sue to stop abuse. A version was lined up for the Senate's defence bill in a package of amendments. When I graded this row, that bill had been stalled since the summer. On 15 September Senators Hawley and Cruz came out against any antitrust exemption for AI. The safe harbour is the waiver step two needs. The checker is public and unpaid by the labs, but there is no line. Any group of firms qualifies for any delay tied to a named risk, so it can carry a brake or a cartel. Sources: bill text; Semafor, 16 Sep; Hawley, 15 Sep.

A standards body on the FINRA model. On 14 July Demis Hassabis proposed an industry-funded, government-backed body, modelled on the self-regulator for American brokers. It would test frontier models for dangerous cyber, biological and deception capabilities up to 30 days before release. It would be voluntary at first. Once the tests prove themselves, it would be required for any frontier model sold in the US. On 15 September OpenAI said it had been talking with Anthropic and Google about working together on safety ever since that proposal. The line is right if it stays at frontier models. The checker is the problem. Standards written by the companies they bind, and checked by staff those companies pay, are a closed loop. They stay one unless a public supervisor holds the power over the body that the SEC holds over FINRA. Hassabis calls it federally overseen but does not say by whom or with what powers. Sources: Hassabis, 14 Jul; CNBC, 15 Sep.

A federal liability shield. Labs have asked Washington to limit what they owe when their models cause harm, according to reports. On 15 September the Treasury Secretary told a House committee that "the best way to guarantee safety is that the creators are liable for what they build and generate". On 21 September he said the government would not be a liability shield. No federal text exists. The nearest is Illinois SB 3444, which stalled in May. It would have spared developers of models above 10^26 operations from liability for catastrophes (a hundred people killed or badly hurt, or a billion dollars of damage). The shield would have applied if they had published safety reports and had not acted intentionally or recklessly. A shield stops nothing a model could do. It leaves the cost of harm with whoever is harmed, which is why What Writing Did calls it an accountability sink written into law. I oppose one on its own. At most it could be part of the price of a real brake, one with a line and a checker the labs don't pay. Even then it is paid for with the public's right to sue. It becomes a moat if only those who sign the incumbents' standards qualify. Sources: FedScoop, 15 Sep; Fortune, 22 Sep; Illinois SB 3444.

Ban Artificial Superintelligence Act. It would ban superintelligence outright. Every model trained with 10^25 operations or more would go under a mandatory pause. There'd be no further training and no new releases until a new Department of Artificial Intelligence is staffed and has written its rules. Building or distributing such a model would need the department's charter. Deploying, releasing, importing or transferring one would need its approval, under section 9(d). Its list of banned capabilities, such as resisting shutdown or breaking into systems, is written the way a brake is. On the test, it stops a capability, so it is a brake. A startup training below the line faces no pause, only the capability bans that bind everyone, so it passes the second question. The checker is public and the labs don't pay it, which is the part worth keeping. But it fails the fourth question. Section 15 tells the government to seek international agreements, but the pause doesn't wait for them. It would stop qualifying runs at home and none abroad, and keep foreign models out with an import wall. That assumes the frontier can be kept at home, while open-weight models trail it by months and arrive as downloads. So the frontier would move to whoever didn't sign. I think section 8 bars only unreleased models. If so, models already released stay on sale while nobody can release a new one. Sections 9(d) and 12(b) require approval to deploy and a charter to distribute, and they may reach them too. The text doesn't settle it. Its line is reset each year to hold the capability of a 10^25 run constant. So the line falls as training gets more efficient, and covers more of the industry. Sources: press release; bill text.

Senate frontier AI draft. Senators Thune, Cruz and Klobuchar have been negotiating a bill since the summer. It would put a duty of care on developers of the most advanced models. As reported, the companies would test their own models and the Commerce Secretary would approve release. Senator Cantwell reportedly wants the national laboratories to do the testing. The draft would reportedly also preempt state safety laws. On 16 September a coalition of safety groups wrote to reject it. No text is public. If the labs grade themselves and the government reads the grades, it is the closed loop the test warns about. If the national laboratories test, the checker is public and the labs don't pay it. Sources: Nextgov; Washington Post, 16 Sep.

FRONTIER Act. Anyone who trains a model above 10^26 operations must publish a transparency report at or before release. The bill counts making a model available for copying or modification as release, so open weights are covered. The heavier duties are a published safety framework, a yearly third-party audit and registration. They fall only on developers with more than $50 million in revenue that have spent at least $1 billion on AI over three years. That's close to the entity trigger Ball and Ramakrishnan proposed. The very largest must also keep a licensed verification organisation, which reports to the government and the lab at the same time. The line can only be raised. The labs pay their auditors but may not tie the fee to the result. That's better than most proposals, and still short of a checker the labs don't pay. The Commerce Secretary could suspend a model that presents an imminent catastrophic risk. Section 9 of the introduced text would also bar states from adopting or enforcing new obligations on developers for frontier risk transparency, third-party audits and incident reporting. That would preempt the core of California's SB 53 and New York's RAISE Act, both graded brakes here, though the bill names neither. Sources: bill text; Obernolte and Trahan, 23 Jul.

California SB 53. A frontier model is one trained with more than 10^26 operations. Every developer of one publishes a transparency report at release, and reports critical safety incidents to the state within 15 days. Developers with more than $500 million in revenue also publish a safety framework. They send the state summaries of their catastrophic-risk assessments too. On the test, it stops something a model could do, a frontier model shipping in silence, so it is a brake. It passes the second question. A startup training a small model faces nothing, and a report is a sliver of the cost of a run that size. It is weak on the third. The labs report to the state and no outside audit is required, so they report on themselves. The Attorney General enforces it, with fines of up to $1 million a violation. Source: the act.

New York's RAISE Act. Signed in December 2025, it was rewritten in March 2026 to match SB 53. It has the same 10^26 line for frontier models and the same $500 million revenue line for the heaviest duties. Incidents are reported within 72 hours. On 21 September the governor appointed the official who will put the law into effect at a new state office. The governor also said developers will register from November and must comply from January 2027, and floated "kill switches" if they prove feasible, which is too early to grade. On 10 November the state said the office had launched a portal where large frontier developers register, with a deadline of 1 January 2027. The same day it launched a research centre at SUNY Binghamton that can, at the state's request, provide assessments to inform the oversight of frontier developers. Rechecked on 11 November, and the verdict stays. The difference from California is who pays the checker. The largest developers fund the office through assessments, but the state runs it. That's close to the levy the essay suggests for evaluators. Universities doing academic research are exempt. Sources: Governor Hochul, 21 Sep; New York DFS, 10 Nov; Morrison Foerster on the amendment; Bloomberg, 21 Sep.

EU AI Act, models with systemic risk. A general-purpose model trained with more than 10^25 operations is presumed to carry systemic risk. Its provider must evaluate it, test it adversarially, assess and reduce the risks, report serious incidents and secure it. These duties have applied since August 2025. The Commission's AI Office has been able to enforce them since 2 August 2026. The omnibus that took effect on 27 July 2026 delayed the high-risk rules and left these duties as they were. The open-source exemption does not apply at this tier, so released weights are bound the same as closed ones. The checker is a public office the labs don't pay. The Commission must adjust the line as the technology changes. Two cautions. Lighter duties (documentation, a copyright policy and a summary of the training data) start at about 10^23 operations for any general-purpose model. So a startup training a small model is not quite free of obligations, though those duties serve transparency and copyright rather than safety. And I know of no public estimate of what compliance costs at the 10^25 line. If it turns out to be a large fixed sum, this becomes a brake and a moat. Sources: the AI Act; Commission guidelines; Regulation (EU) 2026/1744.

Entity-based regulation. Dean Ball and Ketan Ramakrishnan proposed regulating the companies at the frontier rather than their models, the way banks are regulated. Their example trigger is more than $1 billion a year in AI research spending. That exempts the small the way a revenue floor does. A rule on entities is a brake when the entities are defined by what they spend or train. It's a moat when they are defined by who got there first. This one is defined by spend. On its own, a spending trigger can miss a small lab with a frontier-scale run. That's what a compute line is for, and the FRONTIER Act now uses both. Source: Carnegie Endowment, July 2025.

US export controls on AI chips. Controls like these work as a wall at the border, and walls get routed around. This year's record fits. In January the Commerce Department moved sales of H200-class chips to China from a presumption of denial to case-by-case review. At the end of May it said licences are needed to ship to Chinese-headquartered companies wherever they are. By late August it was drafting a rule on remote access. That followed reports that Chinese firms were renting restricted chips in data centres in Southeast Asia. On the test, a wall stops a country, not a capability or a company. It does nothing about open-weight models once they are released. The essay would build a window instead, through compute visibility. A government could then say how many frontier-scale runs happened last quarter, and where. Sources: Federal Register, 15 Jan; Al Jazeera, 1 Jun; Tech Times, 29 Aug.

Bans on open-weight models. In July the administration was reported to be weighing a ban on Chinese open-weight models. That was after the White House accused a Chinese lab of distilling an American model. In July, too, Hugging Face, Meta, Microsoft, Mistral, Nvidia and others wrote against broad restrictions. Nothing is written down. The essay's verdict covers bans in general. A ban buys some protection against misuse, because the person who downloads the weights is the one no rule can see. It pays for it in concentration, by leaving the frontier to the few who can train it. I rank concentration first, so I would refuse the trade. A ban aimed only at one country's models is a different rule, and too early to grade. Sources: TechCrunch, 24 Jul; Lawfare.

A rival test

On 16 September ComplianceHub published a "capture test" for the oversight proposals moving through Congress and the states. It covers licensing, incident reporting, kill switches and a stand-down. It asks who a rule binds, whose costs it raises, and whether the lab asking for it is the lab that benefits. Its core question is whether a rule changes who can build, or only what builders must disclose. We agree that licensing is the moat. We disagree in three places. It counts who is asking. I don't, because that charge can be made of any rule, and the text is what binds. It treats a compute or capability threshold as the part of a regime that shields incumbents. I think the threshold is a brake's best tool, as long as the line moves with the frontier and the cost scales with the run. The licence attached to it is the moat. And it rates incident reporting that binds every developer as the rule least open to capture. Mine first asks whether a startup training a small model is left alone. The reporting duties in SB 53, RAISE and the FRONTIER Act leave it alone. A rule that binds everyone does not. Source: ComplianceHub, 16 Sep.

How to disagree

Tell me the row and what I got wrong. It might be a fact. Or the text might answer something the row missed, like the line, the cost at the threshold, or the checker and who pays them. I'm on X at @cobuskok. If you're right, I'll regrade the row and change its date. I write with Claude, which Anthropic makes, and three rows grade Anthropic's own proposal. Weigh those with that in mind.

Last updated 11 November 2026.