← Back to feed

Dario Amodei published a 5,000-word policy essay on his personal site on June 10, arriving on Hacker News front page within hours. The post is titled “Policy on the AI Exponential” and is the most detailed public policy statement the Anthropic CEO has made. It comes with two concrete deliverables: a legislative proposal on frontier model testing and a funded framework for job displacement.

The Treebeard problem

Amodei opens with a Lord of the Rings metaphor. AI advances at “lightning pace”; policy moves like Treebeard, a sentient tree who takes a full day just to say hello. The mismatch in timescales is the central problem.

He argues that in the four years since AI became a commercial technology, those advocating for responsible development were stuck advocating for “optionality-preserving” policies — transparency legislation, chip export controls, labor market data collection — because that was all that felt politically achievable. The fundamental risks were clear to people inside the lab; they were not yet visible to policymakers.

That changed.

Mythos as the inflection point

Amodei cites Claude Mythos Preview as the proof that AI is now a matter of global strategic consequence. “Mythos Preview scrambled the global cybersecurity landscape,” he writes. It demonstrated that frontier models can pose real risks to financial sector stability, critical infrastructure, and national security.

The implication he draws is larger than cybersecurity: if a model can be demonstrated to pose strategic risk in one domain, that sets a template. He states he believes biological risks may follow, and that “serious AI autonomy risks may not be far behind.”

Within Anthropic’s public communications, this is unusually direct. The lab has previously been careful to qualify existential risk language. Here Amodei is making it an explicit policy call.

Five policy areas

The essay covers five domains, with specific recommended actions in each:

Regulation and public safety. Amodei argues that the right approach in 2023-24 was transparency — requiring disclosure of safety procedures, tests run, and critical incidents. That phase needs to expand to mandatory pre-release testing for frontier models. Anthropic is releasing a legislative proposal for this.

Macroeconomics and tax policy. He argues that AI’s labor displacement effects will require new economic frameworks — both safety nets and redistribution mechanisms. Anthropic is releasing a policy framework for job displacement with stated “substantial financial backing.”

Scientific innovation. AI should be steered toward applied research in biology, medicine, and materials science. He doesn’t detail specific proposals here but frames it as a deployment priority, not just a research curiosity.

Balance of power between state and society. This is the section with the most tension. Amodei argues that AI creates risks of excessive state power concentration — both in authoritarian governments and potentially in democratic ones. He advocates for structural safeguards. He does not name countries but the subtext is clear.

Geopolitics. Amodei calls for coordinated international frameworks — more than bilateral agreements, closer to the nuclear non-proliferation model — for frontier AI development. This section is the least specific and the most aspirational.

What Anthropic is committing to

Two concrete deliverables ship alongside the essay:

1. A legislative proposal on frontier model testing — the idea that governments should have the ability to evaluate models before they are released to the public, similar to pharmaceutical approval logic but designed for AI’s speed.

2. A policy framework for job displacement, with “substantial financial backing” from Anthropic. The size is not specified.

Amodei describes these as “first steps to signal our seriousness.” He explicitly frames them as incomplete and expects to do more.

Why this matters

Anthropic is the most-watched AI safety lab and its CEO has previously been cautious about specific legislative proposals. The move from “here’s what we think is risky” to “here’s draft legislation and a funded program” represents a meaningful shift in posture. It comes at a moment when the lab’s valuation is approaching $1 trillion and it holds multi-year compute deals with Google, Amazon, and SpaceX. The lab has both the credibility and the financial weight to move policy in ways it could not two years ago.

The legislative proposal on frontier model testing is also directly relevant to competitive dynamics. Mandatory pre-release testing, if enacted, would apply to every frontier lab operating in the US — including OpenAI, Google, and xAI. Anthropic has arguably more experience with structured pre-release safety evaluation than any of them.

The essay is available in full at darioamodei.com.