Anthropic’s CEO Wants to Pump the Brakes on AI — Here’s What His Three-Step Plan Actually Means
Anthropic CEO Dario Amodei has published a detailed essay calling for a deliberate slowdown in AI development, proposing a three-step plan that begins with giving third-party evaluators like METR access to Anthropic's models and extends to industry-wide coordination and stronger regulatory frameworks. The move signals a significant shift in how one of AI's most prominent leaders thinks about the relationship between capability development and safety infrastructure.
When the CEO of an AI Company Says “Slow Down,” the World Should Listen
There is something genuinely remarkable about the chief executive of one of the world’s most powerful AI labs publicly calling for the industry to decelerate. That is precisely what Anthropic’s Dario Amodei has done — and the details of how he wants to do it deserve careful unpacking.
As reported by The Verge at https://www.theverge.com/ai-artificial-intelligence/994337/anthropic-ceo-slow-down-ai-development, Amodei has published a wide-ranging essay proposing a structured, three-step plan to “pace the frontier” — a phrase that, stripped of its technical jargon, simply means slowing the rate at which companies train and deploy increasingly powerful AI models. His core argument: the industry needs to pause long enough to build proper safeguards and give regulators the breathing room to actually evaluate what these systems can do.
What “Pacing the Frontier” Really Means
The phrase “pacing the frontier” sounds polished and boardroom-safe, but its implications are significant. In the current landscape, the dominant incentive for every major AI lab — including Anthropic, OpenAI, Google DeepMind, and Meta — is to ship faster, train bigger, and claim capability benchmarks before rivals do. The competitive pressure is immense, and it has historically pushed safety considerations to the side in favour of speed.
Amodei’s proposal essentially asks the industry to voluntarily resist that pressure. He is not calling for an indefinite moratorium or a hard stop to all research. Instead, he is advocating for a deliberate, structured pacing mechanism that creates space for safety infrastructure to catch up with capability development. The distinction matters: this is a call for thoughtful coordination, not a retreat.
Step One: Unilateral Action — Opening the Doors to External Evaluators
The first and most immediately concrete step Amodei has announced is one Anthropic is already taking on its own, without waiting for competitors or governments to follow suit. Anthropic will give third-party evaluators — specifically naming METR, an organisation focused on model evaluation and red-teaming — wide-ranging access to its models.
The goal is to ensure Anthropic’s “adherence to safety practices and commitments,” according to The Verge’s reporting. This is meaningful because it shifts the verification of safety claims away from self-reporting — which has historically been the norm across the industry — toward independent, external scrutiny.
METR and similar evaluators are tasked with probing models for dangerous capabilities, unexpected emergent behaviours, and gaps between what a lab claims its model can or cannot do and what it actually does in practice. Giving such organisations deep access is a significant step, and the fact that Anthropic is doing it unilaterally — before any regulatory requirement demands it — signals a genuine shift in posture, at least within the company itself.
Why External Evaluation Matters So Much
For Indian enterprises and developers increasingly integrating AI tools into their workflows, the question of who is verifying AI safety claims is more than academic. When a company self-certifies that its model is safe, there is no independent check on that claim. Third-party evaluation introduces accountability. It is the difference between a pharmaceutical company self-approving its own drugs and having them reviewed by an independent body like India’s Central Drugs Standard Control Organisation.
The analogy is imperfect — AI regulation is nowhere near as mature as pharmaceutical oversight — but the directional logic is the same. External evaluation creates a layer of trust that self-certification simply cannot.
Step Two: Industry-Wide Coordination
Amodei’s second step moves beyond what any single company can do alone. It envisions the AI industry coming together collectively — likely with some form of government involvement — to agree on shared pacing standards. The idea would be to establish norms around how quickly frontier models are trained and deployed, with all major labs agreeing to the same cadence.
This is where things get genuinely complicated. Coordinating competitors in a high-stakes technology race is extraordinarily difficult. The competitive dynamics that drive rapid development also create powerful incentives to defect from any voluntary agreement. If one lab slows down and another does not, the slower lab loses market position, talent, and investment. Without enforceable rules, voluntary coordination has a fragile history.
There is also the geopolitical dimension. Any discussion of slowing AI development in the United States or Europe immediately raises questions about what happens in China, where labs like DeepSeek are advancing rapidly and are under no obligation to follow Western industry norms. Amodei’s essay, as summarised by The Verge, does not appear to fully resolve this tension — it is a proposal in progress, not a complete solution.
Step Three: Regulatory Frameworks That Can Actually Evaluate Models
The third step in Amodei’s plan involves building regulatory capacity — giving governments and oversight bodies the tools, expertise, and access they need to meaningfully evaluate powerful AI systems before they are widely deployed.
This is arguably the most ambitious and long-term element of the plan. Most regulatory bodies around the world, including in India where the government has been developing its AI governance framework, currently lack the technical expertise to independently assess frontier AI models. Regulation often lags technology by years or even decades, and AI development has been moving so quickly that the gap between what regulators understand and what the technology can do has grown to a chasm.
Building that capacity takes time, investment, and sustained political will. Amodei’s proposal implicitly acknowledges that slowing down development is, in part, a gift of time to regulators — a window in which they can catch up rather than perpetually chase a moving target.
A Self-Interested Virtue Signalling, or Something Genuine?
The obvious question is whether Amodei’s call for restraint is sincere or whether it serves Anthropic’s competitive interests. Cynics will note that Anthropic, which has built its brand around being the “safety-focused” AI lab, benefits enormously from a regulatory environment that emphasises safety checks — particularly if those checks create barriers that are harder for newer or less well-resourced competitors to clear.
There is also the question of timing. Anthropic is not the fastest-moving lab in the world right now. Calling for the industry to slow down when you are not currently leading the pace race is a different gesture than calling for it when you are setting the pace.
That said, dismissing the proposal entirely as self-serving would be too convenient. Amodei has been one of the more consistent voices in the AI industry raising genuine safety concerns over the years. The fact that Anthropic is taking the unilateral step of opening its models to third-party evaluators without waiting for regulatory mandates suggests at least some genuine commitment to the underlying principle.
What This Means for the Broader AI Ecosystem
For developers, enterprises, and policymakers in India and globally, Amodei’s essay represents a significant moment in the ongoing conversation about AI governance. It is not the final word — it is the opening of a negotiation.
The core tension it surfaces is real and unresolved: how do you maintain meaningful safety oversight over technology that is advancing faster than the institutions designed to govern it? Slowing development is one answer. Building better, faster evaluation tools is another. Most likely, the solution will involve elements of both.
What is clear from Amodei’s proposal, as covered by The Verge, is that even the people building frontier AI systems are increasingly uncomfortable with the pace at which they are doing so. That discomfort, expressed publicly and with a concrete plan attached, is worth taking seriously — regardless of what you think of the motivations behind it.
The Bottom Line
Dario Amodei’s three-step plan — unilateral third-party access, industry-wide coordination, and stronger regulatory frameworks — is the most structured call for AI pacing to come from a sitting CEO of a major AI lab. It does not resolve every tension, and it leaves significant questions unanswered about international competition and enforcement. But it shifts the Overton window on what responsible AI development can look like, and it gives regulators, enterprises, and civil society a concrete framework to push back on, refine, or build upon.
The brakes may not yet be applied to the whole vehicle. But someone with their hand on the wheel is finally asking for them.
