Published in September 2026, Dario Amodei’s We Must Pace the Frontier argues that the advance of frontier AI model capabilities should be deliberately slowed so that safety, alignment, evaluation and operational controls have time to keep up. The essay does not call for an end to model training. Its idea of “pacing” is to reduce the relative speed of capability progress so safeguards can catch up.
The essay cites two motivations: faster use of AI to build the next generation of AI — described by the author as the beginning of recursive self-improvement — and what he calls the “OpenAI-Hugging Face incident.” Those passages should be read as the author’s diagnosis and forecast, not as public law or an independent regulatory finding.
The most important point for interpreting the document is to separate three different statuses: there is a voluntary commitment announced by Anthropic, proposals that would depend on coordination among companies and governments, and, finally, international objectives that would be much harder to execute. Treating all three layers as if they were already binding policy overstates what the essay actually proposes.
The short answer: three layers of governance
| Proposal | Responsible actor | Mechanism | Dependencies | Limits |
|---|---|---|---|---|
| Embedded external evaluators | Frontier AI labs | Ongoing, employee-like access to verify practices, report incidents and assess models, training and processes | Evaluator independence, real access to systems and information, confidentiality and publication rules | Initially voluntary; the essay does not define a universal audit standard or sanctioning power |
| Coordination among labs in democracies | Companies, governments and potentially industry groups | Common standards, capability-and-safety “checkpoints,” regulation and voluntary agreements | Shared metrics, government mediation, compatibility with competition law and participation by a critical mass of companies | The concrete triggers are still examples; the proposal may raise antitrust, regulatory-capture and competition disputes |
| Global coordination | The United States, allies and authoritarian governments, especially China in the author’s framing | Graduated agreements, from narrow prohibitions to speed limits on recursive self-improvement | Reliable verification, incentives to comply and the ability to detect secret models or programs | The more ambitious the agreement, the greater the defection risk and the harder the enforcement |
1. Embedded external evaluators: the most concrete part
The first layer is the only one Anthropic says it is unilaterally committing to in the essay itself. The proposal is to give an external team ongoing access similar to that of employees who perform risk assessments. The goal is not merely to test a finished model before release, but to observe training, deployment, operations, safeguards and incidents throughout the process.
How that access would work
- The text envisions a physical and operational presence, with desks in the office, access badges and company laptops.
- Evaluators would have access to workspaces, tools and permissions broadly comparable to internal risk-assessment teams, with exceptions required by law, contracts and the privacy of customers and partners.
- The contract should allow publication of relevant findings about risks, incidents, practices and even about the access that was granted or denied.
- The company could redact information in narrow sensitive categories, but not hide a conclusion simply because it is unfavorable; reviewers could publicly state if a redaction removed something important to their conclusions.
In practice, this layer tries to solve the verifiability problem: safety commitments have limited value if an outside observer only receives reports selected by the company itself. The essay cites METR as an example of an organization that could perform this kind of role, but it does not present a final auditor or a settled accreditation standard.
Dependencies and limits
The design is stronger than a one-off audit, but it still leaves open questions. Who selects and pays the evaluator? What conflicts of interest would disqualify a firm? What methodology would be common across labs? What happens when the evaluator and the company disagree about the severity of a risk? And who imposes consequences if the company does not follow a recommendation?
For that reason, an “embedded evaluator” is not automatically the same thing as an independent regulator. In the document, the mechanism begins as a corporate commitment and as a prototype for something Amodei wants governments to formalize.
2. Coordination among labs: a safety pact within democracies
The second layer requires industry coordination. Amodei says he prefers regulation that covers all US frontier labs, because a voluntary rule does not cover companies that refuse to participate. At the same time, he argues that legislation can take time, so companies should negotiate voluntary standards in parallel.
The essay recognizes that some forms of coordination among competitors are legally sensitive. It therefore suggests government mediation or a narrow waiver from antitrust restrictions for specific safety discussions. It also mentions the possibility of using industry groups associated with government.
Capability-based checkpoints
The most operational proposal in this layer is to create “checkpoints” tied to what a system can do. The essay’s example is: if a model reaches capability X, it should only proceed alongside evidence of safety properties Y and Z, combining evaluations, interpretability and audits of training environments.
Amodei also says the discussion could include limits on development inputs — such as training compute, the nature of training runs and internal use of AI to improve AI — but notes that those indicators may be easier to game than external model behavior. In other words, the essay offers a control architecture, not a ready-made table of thresholds.
The geopolitical condition in the argument
This second layer is not presented as an abstract safety pact. The author conditions the amount of slowing on the technological lead of democracies over projects associated with the Chinese Communist Party. He advocates chip and equipment controls, action against smuggling and remote access to data centers, a crackdown on unauthorized distillation and stronger protection against model-weight theft.
Those points are strategic premises of the author’s argument. They help explain the document’s logic, but they should not be confused with neutral technical requirements or rules already approved for all labs.
3. International coordination: from narrow agreements to broad pacing
The third layer attempts to move “pacing” to the international level. The essay itself treats this as the hardest part, because any broad agreement would depend on verifying that all sides actually comply with limits and do not keep secret models capable of changing the military or economic balance.
| Level | What would be negotiated | Difficulty described in the essay |
|---|---|---|
| 1 | Ban narrow and clearly dangerous uses, such as using AI to produce biological weapons | Presented as relatively plausible because the harm is against every side’s interest |
| 2 | Test models before release for acute risks in cybersecurity, biology and alignment, potentially through a global standards body | Creating standards may be feasible; verifying secret programs would be much harder |
| 3 | Apply a “speed limit” to recursive self-improvement | Difficult but described as being on the edge of possible, because it would reduce speed without giving up all strategic advantage |
| 4 | Broad pacing or a substantial pause in development | The author supports discussing it but considers it unlikely in the near term because incentives to evade the agreement would be large |
What “pacing” is supposed to buy, according to the essay
The core argument is not to slow down for its own sake. Amodei argues that even one or two additional years before reaching critical capabilities could be used in four areas: operational excellence, alignment, interpretability, and stronger testing and evaluation.
The reasoning is hypothetical: the essay does not demonstrate that coordinated slowing will necessarily create that amount of time or that the time will be sufficient. It argues that the interval could reduce risk if it is actually converted into technical and organizational improvements. That condition — using the time well — is an essential part of the proposal.
What the document does not establish
- It does not create a law, treaty or regulator with enforcement powers.
- It does not set a universal schedule for reducing the speed of model training.
- It does not define mandatory numerical thresholds for capability, compute or risk.
- It does not define who would certify every external evaluator or how disputes between auditor and company would be resolved.
- It does not turn the four international levels into existing agreements; they are graduated options for future negotiation.
- It does not simply propose stopping all AI research. The text distinguishes “pacing” from a general halt to technical progress.
Where the limits and controversy lie
Coverage after the essay recorded support from some industry leaders, but also skepticism about a model in which companies select their own evaluators and seek legal room to coordinate standards among competitors. Critics quoted by The Guardian raised concerns about self-regulation, regulatory capture and possible competition effects.
Those criticisms do not erase the mechanism described, but they identify its weak point: credibility depends on real independence, sufficient transparency, comparable criteria and consequences when commitments are broken. Internationally, the problem is even larger, because the proposal itself acknowledges that verification and defection risk constrain what can be negotiated.
How to read the document without confusing opinion with binding rules
A useful documentary reading separates three mental columns. Commitment: Anthropic says it will move ahead with embedded external evaluators. Proposal: Amodei asks other labs and governments to formalize standards, checkpoints and coordination. Aspiration: the essay imagines increasingly ambitious international agreements while admitting that the highest levels may be impractical in the near term.
This framing avoids two opposite mistakes. The first is treating the essay as though it had already created new rules for the industry. The second is reducing it to an abstract call for “more safety.” There are relatively concrete mechanisms — especially ongoing evaluator access and capability-based checkpoints — but their effectiveness depends on governance, incentives, metrics and enforcement that the document does not yet close.
In one sentence
We Must Pace the Frontier proposes replacing internal safety promises with a progressively more verifiable system: first external evaluators inside labs, then coordinated standards among companies and democratic governments, and finally graduated international agreements. The first step is an announced corporate commitment; the others remain policy and diplomatic proposals, not binding rules.
Sources: Dario Amodei — We Must Pace the Frontier; The Guardian — AI CEOs say they need to slow the pace of development. But will they?.
