In June 2026, Dario Amodei published a long policy essay: Policy on the AI Exponential.
This is not simply an essay about AI safety, nor simply one about regulation. Its real point is a timing mismatch: AI capability is moving along an exponential curve, while the policy system still moves at traditional speed.
Using his metaphor, AI is like the hobbits in The Lord of the Rings urgently asking for help, while political institutions are like slow-moving Ents. The problem is not that the Ents never wake up; it is that they wake up too slowly. By the time legislation, regulation, and international coordination slowly form, AI may already have moved from “an interesting tool” to “a whole country of geniuses in a data center.”
The essay is worth reading not because it gives every answer, but because it lays out the policy issues Anthropic now cares about most: pre-deployment testing, labor disruption, scientific regulation, civil liberties, and an alliance of democratic countries.
The Core Argument
Dario’s main line is clear: in the past few years, AI risks were not yet concrete enough, so the most realistic policies were transparency, disclosure, chip export controls, and data collection on labor impacts.
Those policies were meant to preserve options for future response.
But he argues that the situation has changed. Frontier models have already become capable enough in areas such as cybersecurity to make AI a tool of national strategy. Future risks may also include biological risks, loss of control over autonomous systems, and automated R&D that accelerates these risks.
So policy can no longer stop at “let’s wait and see.” He argues for a stronger constraint phase: frontier models above a certain compute threshold should undergo mandatory third-party testing; if a model presents unacceptable risk in specific areas, the government should have the power to block or delay deployment.
This is not treating AI like ordinary software. It is closer to how we treat aircraft, medicine, or cars: technologies that can bring huge benefits, but can also cause large-scale harm if designed or operated badly.
Part One: Regulation Must Move From Transparency To Pre-Deployment Testing
The first part discusses regulation and public safety.
He admits that regulating too early can easily go wrong. In 2023 and 2024, the direction of the risks could be seen, but the risk shapes, testing methods, and mitigations were not yet fully clear. Legislation at that point might have created many low-value compliance requirements while missing the risks that actually mattered.
That is why Anthropic supported transparency legislation at the time: requiring model developers to disclose safety processes, testing methods, and major safety incidents.
Now he believes transparency alone is not enough.
His proposed direction includes:
- Models above a compute threshold must be tested by qualified third parties across four risk areas: cybersecurity, biological weapons, loss of control over AI systems, and automated R&D that could accelerate these risks.
- If third-party evaluation finds unacceptable risk, the government should have the power to prevent or deter deployment.
- Testing could be done by a government agency similar to the FAA, or by private evaluators authorized and supervised by the government.
- Companies developing frontier models must protect model weights, conduct regular red-team tests and penetration tests, and work with the government to defend against major threat actors.
- Safety incidents in the four critical risk categories must be reported promptly.
The key phrase here is “specific risks.” He is not arguing for unlimited government takeover of AI. He is limiting authority to the four most serious risk categories and asking for safeguards against political favoritism and arbitrary discretion.
Part Two: Employment Is Not A PR Problem
The second part discusses macroeconomics and tax policy.
Dario’s view has two sides.
One side is strongly optimistic: if AI can perform most cognitive tasks far beyond humans, it may create extremely fast economic growth through science, technology, and operational efficiency. AI can also help build better AI, which further amplifies growth.
The other side is strongly concerned: for the same reason, because AI can replace broad cognitive capabilities, it may disrupt labor markets faster and more deeply than past technologies.
He says it directly: we may enter a world of “super growth, super inequality.” At that point, the biggest challenge is no longer how to stimulate growth, but how to ensure everyone shares in the gains.
Two points in this section stand out.
First, he stresses that permanent unemployment is not the goal, but a risk to avoid as much as possible. Anthropic does not want companies to use AI only to cut jobs; it wants companies to find new revenue and new use cases so existing employees can do more.
Second, employment policy cannot only solve the problem of “giving people money.” It also has to deal with meaning, purpose, and agency. Economic support matters, but how people find value in a world where AI is stronger than they are is the deeper question.
His policy suggestions include:
- Build better labor-market measurement systems to track AI’s real impact on jobs.
- Use wage insurance, retention tax incentives, training subsidies, and labor-matching infrastructure to slow displacement.
- If labor demand falls over the long term, consider long-term income support, such as universal basic income, capital gains taxes, and universal capital accounts.
He also mentions the debate over data centers and electricity prices: AI companies should bear the cost of electricity price increases, but public resentment toward data centers is also an outlet for anxiety about AI’s economic impact.
Part Three: AI’s Positive Benefits Can Also Be Blocked By Old Regulation
The third part discusses scientific innovation, especially biomedicine.
Here his view is somewhat reversed from the first part.
For AI itself, he worries that regulation is not moving fast enough. For downstream technologies accelerated by AI, he worries that regulation is too slow and too old to absorb the pace of innovation.
In drug development, for example, AI could bring changes such as:
- A large increase in the number of new drug candidates entering regulatory pipelines.
- Better drug efficacy and safety.
- Candidate drugs for diseases that previously could not be treated.
- Rapid creation of new treatment modalities.
But existing drug regulation is built on a relatively pessimistic assumption: most candidate drugs fail, and even effective drugs may have serious safety issues. Agencies such as the FDA and EMA therefore usually take years to complete approval.
If AI causes a flood of candidate drugs, the old process may become clogged.
His suggestion is not to lower safety standards, but to build new ones in advance: which clinical steps can be replaced by AI simulation or analysis, and under what conditions regulators should accept AI-based PD/PK modeling, toxicology prediction, dose selection, biomarker validation, synthetic control arms, and surrogate endpoints.
This section matters. It reminds us that AI policy is not only about “how to restrict AI”; it is also about “how not to let old institutions slow down the benefits AI creates.”
Part Four: State Power And Civil Liberties Will Be Amplified By AI
The fourth part discusses the state, civil liberties, and checks on power.
Dario’s concern is that strong AI in the wrong hands could become a tool of authoritarianism. Existing legal and constitutional protections may not be ready for the sudden power jump AI could create.
He gives several risk examples:
- Fully autonomous weapons might obey illegal orders, allowing governments to bypass the professional constraints of human officers.
- Surveillance-oriented AI could analyze public and private data at scale, inferring the deepest details of citizens’ lives.
- Governments or companies could gain excessive power through AI, even bypassing democratic oversight.
His policy suggestions include:
- Establish reliable accountability rules for fully autonomous weapons, ensuring they respond to court orders, legislation, and senior human oversight rather than blindly obeying commands.
- Ban fully autonomous weapons in domestic law enforcement.
- Close data-broker and large-scale data purchase loopholes so private-company data cannot be used for domestic surveillance and enforcement.
- When the government takes adverse action against an individual or organization, the affected party should have access to AI advice of at least equal capability, preventing the government from holding a one-sided AI advantage.
The logic here is that AI does not only change productivity. It also changes the distribution of power. If policy only looks at economics and safety, it will miss freedom and institutional constraints.
Part Five: Democratic Countries Need To Rebuild Alliances Around AI
The fifth part discusses geopolitics.
Dario explicitly opposes treating AI merely as a trade-policy tool. He sees AI as a strategic variable on the level of nuclear weapons, and perhaps even more important.
If strong AI really can become something close to “a country of geniuses in a data center,” it will become a major source of national military and economic capability. A country with strong AI could form an overwhelming advantage over one that is three years behind.
Therefore, he argues that democratic countries should build a global alliance around creating AI according to shared values.
This alliance should do several things:
- Manage the AI supply chain, sharing chips and semiconductor manufacturing equipment within a trusted alliance while restricting adversaries’ access.
- Coordinate international regulatory standards on biological, cybersecurity, and autonomy risks.
- Share AI economic benefits and beneficial deployment experience, helping developing countries gain from AI.
- Defend one another in cyber defense, drones, manufacturing, secure compute, AI-driven R&D, and intelligence analysis.
- Reject AI-driven high-tech authoritarian repression.
- Coordinate policy on employment and macroeconomic disruption.
He hopes this alliance starts with democracies that share similar values, then gradually attracts more countries. The ideal outcome is that the whole world eventually joins; if not, at least democratic countries should be in the strongest position to deter and outperform regimes that continue down a repressive path.
What Makes This Essay Valuable
To me, the most valuable part of the essay is not any single policy proposal, but the way it divides AI policy into five layers:
- How to govern the model itself.
- How to govern labor and wealth distribution.
- How to approve scientific results accelerated by AI.
- How to constrain state and corporate power.
- How democratic countries organize alliances in global competition.
This is closer to the real problem than simply arguing “open source or closed source” or “regulation or no regulation.”
The difficulty with AI is not that it disrupts only one department. It is simultaneously like software, infrastructure, military capability, a scientific accelerator, a labor substitute, and an amplifier of state power.
So policy cannot rely on a single line either.
We Should Also See Its Position
This essay comes from the CEO of Anthropic, so it is certainly not a neutral academic paper.
It naturally emphasizes frontier model risks, pre-deployment testing, and government involvement, as well as democratic alliances and supply-chain control. These positions are related to Anthropic’s product position, governance, and identity as a U.S. company.
But that does not mean the essay is not worth reading.
On the contrary, its value comes precisely from that position: a frontier model company is publicly saying that AI is no longer an ordinary technology that “the market will digest on its own,” but an exponential variable that requires the policy system to speed up.
You may disagree with his specific proposals, but it is hard to ignore the judgment: if AI capabilities keep improving rapidly, slow policy meeting fast technology will eventually cause problems.
Summary
The core of Policy on the AI Exponential is to move AI policy from “whether regulation is necessary” to “how the policy system can catch up with the exponential curve.”
Dario Amodei’s answer is: frontier models need pre-deployment testing; labor impacts need early measurement and cushioning; positive applications such as biomedicine need regulatory pathways to be reformed; civil liberties need protection against AI-amplified state and corporate power; democratic countries need alliances around supply chains, safety standards, and values.
This package is not complete, and it will certainly be controversial.
But it raises a question that is getting harder to avoid: when AI can travel in one year the distance previous technologies took ten years to cover, can policy keep moving at the old speed?
References: