It's not every day a CEO asks the government for the power to shut down his own industry's products. But on June 10, 2026 — just one day after Anthropic launched Claude Fable 5, its most capable model yet — CEO Dario Amodei did exactly that. In a sweeping essay titled "Policy on the AI Exponential," he called for mandatory third-party testing of frontier AI and government power to block models deemed unsafe.
It's the most detailed regulatory blueprint any major AI lab has published — and Anthropic backed it with $350 million in new funding. Here's what Amodei is actually proposing, and why the timing matters.
The News in Brief
- What: An essay, "Policy on the AI Exponential," laying out a five-part AI governance plan.
- The headline ask: mandatory third-party testing of frontier models, with government power to block or reverse deployment if a model fails.
- The shift: Anthropic moving from advocating transparency to backing binding regulation.
- The money: $350 million — a $200M Economic Futures Research Fund and a $150M national fellowship program.
- The timing: published one day after the launch of Claude Fable 5.
The Big Shift: From Transparency to Binding Rules
For years, Anthropic's public position on regulation leaned toward transparency — labs disclosing how they test models and what safeguards they use, largely voluntarily. Amodei's essay marks a clear escalation: he now argues that transparency isn't enough, and that governments need real, enforceable authority over the most powerful systems.
The framing is right there in the title. Amodei's core belief is that AI capability is on an exponential curve — improving so fast that policy built for last year's models will always be a step behind. His answer is to regulate the frontier: the small number of most-capable, highest-compute models where the biggest risks concentrate, while leaving the vast majority of everyday AI untouched.
FAA-Style Testing for Frontier Models
The centerpiece of the blueprint is a system Amodei likens to how the FAA certifies aircraft before they're allowed to fly. Frontier models above a certain compute threshold would have to pass independent, third-party testing before shipping — and if a model fails, the government could block or reverse its deployment.
Crucially, the testing is scoped to a specific, narrow set of catastrophic risks rather than open-ended judgment calls:
| Risk area | What it covers |
|---|---|
| Cybersecurity | Whether a model could enable serious cyberattacks |
| Biological weapons | Whether it could help create bioweapons or other CBRN threats |
| Loss of control | Whether the model could act in ways its operators can't reliably control |
| Automated R&D | Whether it could accelerate the other three risks by speeding AI development itself |
To address fears of abuse, Amodei says the framework should include protections against political favoritism — so the power to block a model can't simply be turned into a tool for punishing rivals or rewarding allies.
The Five Pillars
Safety testing is just one part. The essay is structured around five pillars, which together amount to a comprehensive theory of how society should adapt to powerful AI:
"The rules have to move as fast as the technology. You can't govern an exponential with a static rulebook."
| Pillar | The proposal |
|---|---|
| Safety regulation | Mandatory third-party testing of frontier models, with deployment blocks for failures |
| Economics & taxes | Wage insurance, retention tax incentives, training grants; UBI if displacement is lasting |
| Science acceleration | Using AI to speed up scientific and medical research for public benefit |
| Civil liberties | Ban fully autonomous weapons in domestic law enforcement; close the data-broker surveillance loophole |
| Geopolitics | Keeping democracies ahead of authoritarian rivals in frontier AI |
On civil liberties, two asks stand out: Amodei wants fully autonomous weapons banned from domestic law enforcement, and he urges Congress to close the "data broker loophole" that lets agencies buy bulk surveillance data they'd otherwise need a warrant to obtain.
$350 Million for the Jobs Problem
Unusually for a policy essay, this one came with a checkbook. Anthropic pledged $350 million to study and cushion AI's impact on work:
- $200 million Economic Futures Research Fund — to research how AI is reshaping the economy and labor market.
- $150 million national fellowship program — aimed at early-career Americans entering an AI-transformed workforce.
The economics pillar goes further on paper, floating wage insurance, retention tax incentives, and workforce training grants — and, if AI-driven job displacement proves enduring, universal basic income funded through company or capital-gains taxes. It's a striking acknowledgment from an AI leader that the technology his company builds could disrupt millions of jobs.
Why Now — and Why the Skeptics Push Back
The one-day gap between shipping Fable 5 and demanding oversight isn't an accident — it's the argument. Amodei is effectively saying: we just released something this powerful, and you should be able to stop the next one if it's dangerous. It's a bid to set the terms of regulation before governments write their own, less informed, versions.
Critics see it differently. Some argue that incumbents calling for licensing-style rules risk building a moat — heavy compliance burdens that big, well-funded labs can absorb but smaller competitors and open-source projects cannot. Others question whether any government can test for risks like "loss of control" when the science of measuring them is still immature. And the harder economics ideas, like UBI, face long political odds.
Either way, the essay reframes the debate. It lands at a moment when AI is getting cheaper and more pervasive by the month — a dynamic we explored in our piece on the 2026 AI price war — making the question of who can pull the brakes more urgent, not less.
Frequently Asked Questions
What is Dario Amodei's 'Policy on the AI Exponential'?
It is an essay published on June 10, 2026 by Anthropic CEO Dario Amodei, one day after Anthropic launched Claude Fable 5. In it, Amodei lays out a five-part policy blueprint for governing rapidly advancing AI, including mandatory third-party safety testing of frontier models and government power to block or reverse the deployment of models that fail. It marks Anthropic's shift from advocating transparency to backing binding regulation.
What does Amodei want governments to be able to do?
Amodei proposes FAA-style mandatory testing before the most powerful frontier models ship. Models above a certain compute threshold would undergo independent third-party testing in four risk areas — cybersecurity, biological weapons, loss of control, and automated R&D that could accelerate the other three. If a model fails, the government would have the power to block or reverse its deployment. The proposal is scoped to those four risks and includes protections against political favoritism.
What are the five pillars of Amodei's AI blueprint?
The essay covers five areas: safety regulation (mandatory frontier-model testing), economics and taxes (cushioning AI's impact on jobs), science acceleration (using AI to speed up research), civil liberties (limits on surveillance and autonomous weapons), and geopolitics (keeping democracies ahead in AI). Together they form what Amodei frames as a comprehensive "permission layer" for the AI era.
How much money is Anthropic pledging, and for what?
Anthropic pledged $350 million to help manage AI's labor-market effects: a $200 million Economic Futures Research Fund to study AI's impact on the economy, and a $150 million national fellowship program for early-career Americans. Amodei's essay also floats wage insurance, retention tax incentives, workforce training grants, and — if job displacement proves enduring — universal basic income funded by company or capital gains taxes.
Why is it notable that Amodei published this right after Claude Fable 5?
The timing underscores Amodei's central argument: AI capabilities are advancing exponentially, so the rules need to advance with them. Releasing a powerful new flagship model one day, then calling for binding government oversight the next, signals that Anthropic believes the industry cannot rely on voluntary safety commitments alone. It also puts competitive pressure on rivals, since Amodei is effectively inviting regulation of the very products his own company sells.
Final Thoughts
Amodei's blueprint won't become law overnight, and parts of it — especially the tax-and-jobs proposals — face steep political resistance. But as the clearest, most detailed regulation plan yet from inside a leading lab, it sets a reference point the whole debate will now revolve around. When the company shipping frontier models is also drafting the rules to constrain them, regulators, rivals, and the public are all forced to respond.
The throughline of 2026's AI story is acceleration — cheaper models, more powerful systems, bigger bets on the physical world. Amodei's essay is a bet of a different kind: that the guardrails can be built fast enough to keep up. We'll keep covering both sides of it. For more, read our explainers on Claude Fable 5 and the AI price war.