Primary text: darioamodei.com Author: Dario Amodei, CEO of Anthropic. Published June 10, 2026.
"Policy on the AI Exponential" is an essay published June 10, 2026 by Dario Amodei arguing that the evidence of advanced AI's power and risks has become definite enough that policy should move beyond transparency to binding regulation of frontier models. Amodei frames the central problem as a mismatch in timescale between AI's exponential progress and the slow pace of legislation, and argues that early policy actions are "at least a year out of step" with capability gains. The essay was released alongside an Anthropic legislative proposal on frontier-model testing and an economic-policy framework for job displacement, both of which Anthropic said it would back financially.
Core argument
Amodei argues that safety advocates, including Anthropic, had until recently focused on policies that preserve optionality and tee up a faster reaction later — transparency legislation, chip export controls, and data collection on AI's labor effects — because the shape of AI's risks was not yet clear. He contends that the Claude Mythos Preview cybersecurity results, and forthcoming biological and autonomy risks, now make those risks definite, so legislation can be targeted precisely rather than written in the abstract. The essay treats the scaling laws as having "over a decade of empirical evidence" and reasons that if they hold for "a year or two longer," the result is "Powerful AI" — "a country of geniuses in a datacenter." The essay is structured around five policy areas.
1. Regulation and public safety
Amodei argues frontier AI should be regulated on the model of agencies like the Federal Aviation Administration: models should undergo technical testing and auditing, and their release should be blocked or reversed if they do not meet safety standards. The Anthropic proposal he describes includes:
- Models above a compute threshold should undergo mandatory testing by a qualified third party for risk in four specific areas: cybersecurity, biological weapons, loss of control of AI systems, and automated R&D that could accelerate those risks.
- The government should have the power to block or deter deployment of a model determined to present unacceptable risks, with the power scoped to those four risks and protected against political favoritism or arbitrary decisions.
- Third-party evaluation could be performed by a government agency (an FAA analog) or by private organizations authorized and inspected by the government — a "regulatory markets" approach.
- Developers of advanced models must maintain strong security standards protecting model weights, conduct regular red teaming and penetration testing, and work with the government to defend against major threat actors.
- Safety incidents in the four critical areas must be reported promptly.
Amodei contrasts this with the 2023–2024 period, when Anthropic concluded that transparency was the right approach because the form of the risks was unclear, and supported transparency legislation including California SB 53, the New York RAISE Act, and Illinois SB 315. He credits the Trump administration's June 2, 2026 executive order (Promoting Advanced AI Innovation and Security) with moving "incrementally" toward a greater government role, while saying Anthropic's proposal recommends going further. He allows that if the most powerful systems come to resemble "weaponizable nuclear materials" rather than airplanes, more aggressive measures may be needed, but argues against getting ahead of present dangers.
2. Macroeconomics and tax policy
Amodei argues that powerful AI may scramble the usual premise that economic growth is fragile and must be traded off against redistribution, producing a world "stuck on the hypergrowth, hyper-inequality setting." He makes two points emphatically: that enduring job displacement is undesirable and that he warns about it to prompt adaptation, not to bring it about; and that any response must address both economic provision and people's need for meaning and purpose, the latter being a question for society rather than policy. He acknowledges "a decent possibility" that AI causes significant enduring job loss as an "intrinsic property" of a technology that broadly replicates human cognition. Proposed interventions include: expanded government measurement and tracking of AI job displacement (citing Anthropic's Economic Index); pro-employment incentives such as wage insurance, retention tax incentives, workforce-training grants, and employer–employee matching infrastructure; and, if displacement is large and permanent, long-term income support including universal basic income financed through taxes on relevant companies or higher capital-gains taxes, or universal capital accounts. On data centers and energy prices, Amodei argues AI companies should absorb rate increases — citing Anthropic's pledge to do so — and characterizes hostility to data centers as an outlet for broader economic anxiety about AI.
3. Accelerating AI's positive impact
For technologies downstream of AI — biomedicine, energy, materials science — Amodei argues the greater risk is that regulatory systems designed for a slower pace will slow beneficial progress, rather than failing to address risk. Using biomedicine as the illustrative case, he notes the typical FDA or EMA drug-approval pipeline runs 7–8 years and argues agencies should develop standards now for accepting AI simulation and analysis in place of slow experiments — in areas such as AI-based pharmacodynamics/pharmacokinetics modeling, toxicology prediction, dose selection, biomarker validation, synthetic control arms, and surrogate endpoints — and should adopt more flexible accelerated-approval mechanisms.
4. The state and civil liberties
Amodei argues that AI threatens the balance between state power and individual liberty while raising its stakes, and that "powerful AI in the wrong hands could be the ultimate tool of autocracy." He proposes: accountability rules for fully autonomous weapons (and any systems coordinating them) requiring response to constitutional and command oversight rather than blind obedience; a ban on the domestic use of fully autonomous weapons, including in law enforcement; closing the "bulk collection / data broker loophole" that lets data Americans share with private companies be purchased for domestic surveillance; and a public right to AI advice at least as capable as the government's during adverse government action, framed as an extension of the Administrative Procedure Act, due-process protections, or the Sixth Amendment. He adds that companies, not only governments, can capture quasi-state power, citing Anthropic's Long-Term Benefit Trust as one accountability structure and arguing both companies and governments need meaningful checks.
5. Securing leadership by democracies
Amodei argues AI is better understood geopolitically as akin to nuclear weapons than as an instrument of trade policy, and that it is "likely to be the dominant source of military and economic power for any nation." He proposes that democracies form a coalition built around shared values that draws in the rest of the world over time, combining: management of the AI supply chain (coalition members sharing chips and semiconductor manufacturing equipment while denying them to adversaries, with export controls expanded and coordinated, citing the MATCH and OVERWATCH bills); internationally coordinated risk regulation; shared diffusion of AI's economic benefits; mutual defense; rejection of AI-powered repression; and macroeconomic cooperation. The coalition would rest on coordination among sovereign states, starting with aligned democracies and expanding iteratively.
Framing and reception
Amodei rejects the framing that AI's reception is "a PR problem," arguing public concern is a correct response to real risks and constitutes "democratic accountability working as it should." He closes by arguing many of the issues — job displacement, pre-release testing, export controls, energy — have cross-partisan appeal. The essay was covered the same week as a policy statement: it was reported alongside an Anthropic "Advanced AI Framework" (the legislative-testing proposal) and an "Economic Policy Framework" backed by a reported $350 million ($200 million for an Economic Futures Research Fund and $150 million for a national fellowship), with the economic framework floating universal capital accounts and a public fund giving Americans stakes in AI companies (Source: venturebeat.com; politico.com; theinformation.com). Reporting put the testing threshold at models trained above 10²⁵ FLOPs, or built by companies with over $500 million in AI revenue or $1 billion in AI R&D (Source: insideaipolicy.com).
The proposal stands in contrast to the same-week OpenAI plan "Built to benefit everyone" (Altman and Pachocki, June 8, 2026), which emphasizes broad distribution of power and an international coordinating organization over binding national testing and deployment-blocking authority.
Provenance
Single-author essay published on the author's personal site, darioamodei.com, June 10, 2026. The raw file was pulled and authenticity-verified on the canonical host during the June 11, 2026 gap scan, corroborated by Inside AI Policy, VentureBeat, and Politico coverage and the recursive-self-improvement data in "When AI Builds Itself". As an essay advancing the author's policy argument, it is treated as a position, not as factual evidence for the claims it asserts.
Relationships
- depends-on: Scaling Laws; Recursive Self-Improvement (RSI).
- supports: Risk-Based AI Regulation (a binding, capability-threshold variant); Export Controls (AI); AI and Civil Liberties; AI Public Wealth Fund and Government Equity in AI.
- contradicts: (partial) Built to Benefit Everyone: Our Plan (Altman & Pachocki, OpenAI, June 2026) — favors binding national testing and deployment-blocking authority over OpenAI's broad-distribution and international-coordination emphasis.
- related: Anthropic; Dario Amodei; EO — Promoting Advanced AI Innovation and Security (Trump, signed June 2, 2026); The Adolescence of Technology; Machines of Loving Grace.