Anthropic CEO: The government should have the authority to veto high-risk AI, require mandatory testing before it goes live, and the three main proposals conflict with Trump’s deregulatory agenda

Anthropic CEO Dario Amodei Issues Long Statement, Urging Governments to Legislate Mandatory Third-Party Testing for Powerful AI Models, Comparing AI Regulation to FAA Oversight in Commercial Aviation. The Three Main Proposals Cover Deployment Thresholds, Cybersecurity, and Labor Replacement.
(Background recap: Claude has 80% of its code written by itself, Anthropic calls for a "global design brake mechanism"—is it serious?)
(Additional background: Anthropic CEO Dario Amodei: In 6-12 months, Chinese open-source AI models will catch up to Mythos)

Table of Contents

Toggle

• The Three Main Proposals: Thresholds, Weapons, Replacement
• Why Now, the Logic and Limits of the FAA Analogy
• Confronting the Trump Deregulation Pattern

Just yesterday, they released their most advanced general-purpose model, and the next day, they turn around to say AI is too dangerous and should be legislated. This is the signal CEO Dario Amodei of Anthropic sent to the global tech policy community today (11). In his long article "Policy on the AI Exponential," he publicly calls for government intervention for the first time as the company's top leader, advocating for legally binding regulations on the release of "frontier AI models."

His analogy is commercial aviation: airplanes must pass safety reviews by the FAA (Federal Aviation Administration), and AI models should be the same. Amodei said on X:

"Anthropic has long advocated for transparency requirements for frontier AI because the risks weren't yet clear enough for precise regulation. But that is no longer sufficient."

Anthropic has long advocated for transparency requirements for frontier AI, because the risks weren't yet clear enough to regulate precisely. That is no longer sufficient.
— Dario Amodei (@DarioAmodei) June 10, 2026

The Three Main Proposals: Thresholds, Weapons, Replacement

The first key point of the advanced AI framework is setting mandatory third-party testing thresholds: companies training models with over 10^25 FLOPs (floating-point operations, roughly the amount of computation used in training), or with annual AI revenue exceeding $500 million, or R&D spending over $1 billion, must have their models undergo independent audits before release.

The testing focuses on four risk areas: cybersecurity, biological weapons, AI system runaway, and automated R&D that could accelerate these risks. Amodei states: "Frontier AI models are like airplanes—they should be required to pass technical tests and audits; if they do not meet high safety standards, their release should be considered a public safety threat and blocked or withdrawn." Under this framework, governments would be legally empowered to block, delay, or discourage deployment.

The second proposal positions AI as a critical cybersecurity infrastructure issue. Amodei directly cites Anthropic's own model Claude Mythos Preview as an example, which can identify high-severity vulnerabilities in major operating systems, indicating that attack and defense capabilities are advancing in tandem. The framework demands that leading developers protect "model weights" (the core parameters stored after training—if stolen, the entire model can be replicated) from external attackers or insiders, and establish legal channels for reporting "model distillation attacks."

The third proposal, the most politically sensitive, openly acknowledges structural labor displacement. The economic policy framework states that if AI reaches predicted capability levels, it will be a "comprehensive substitute for the workforce," not just a productivity aid. The framework explores scenarios with 5%, 10%, or even more extreme unemployment rates, and advocates for mechanisms like wage insurance, universal basic income (UBI), and sovereign wealth funds.

To support these proposals, Anthropic announced a $350 million investment: $200 million to establish the "Economic Future Research Fund" for pilot public policies, and $150 million for nationwide grants and scholarship programs. Amodei concludes: "The key challenge is not incentivizing growth, but finding ways for everyone to share in the benefits."

Why Now, the Limits and Logic of the FAA Analogy

This framework is released at this moment for two main reasons.

First, the technological pace. Amodei argues that: in the past, because risks were not yet clear, precise legislation was premature; now, with Claude Mythos Preview capable of proactively identifying system vulnerabilities, the "unclear" risk rationale no longer holds. They base legislative arguments on their own model's dangers, which is somewhat convincing, but it also means that as models become more powerful, the urgency for regulation grows.

Second, the boundaries of the FAA analogy. Commercial aviation regulation is straightforward: plane failures have clear physical causes, testing standards are quantifiable, and accident responsibility can be assigned. The "danger" of AI models still heavily depends on evaluators' judgment frameworks, the definition of "frontier" shifts every few months, and the credibility and independence of third-party testing agencies are not yet established. While the FAA analogy is appealing as a rhetorical device, applying a mature industry’s regulatory structure to a rapidly evolving, boundary-shifting technology domain presents many complex challenges—political and technical alike.

Amodei also admits that this framework is a starting point, not the final solution.

Confronting the Trump Deregulation Pattern

Currently, U.S. policy favors deregulation: the Trump administration aimed to let AI industry "grow wildly," using positive competition to beat China, and even actively dismantle local regulatory barriers across states.

Amodei calls for increased regulation, directly opposing this trend. In his concluding remarks, he seeks bipartisan language: "These policy ideas have bipartisan appeal across the political spectrum. The earlier we act, the sooner everyone can share the benefits of AI." However, whether this "common-sense appeal" can translate into legislative momentum remains an open question within the current Washington political landscape.