TL;DR: Anthropic CEO Dario Amodei published "Policy on the AI Exponential" on June 10, 2026, calling for mandatory third-party safety testing of frontier AI (modeled on FAA aviation), pro-employment retention incentives, and UBI funded by an AI company tax if automation becomes permanent. Anthropic committed $200 million to back the research agenda. None of this is law. But when the developer of Claude explicitly endorses binding regulation and an AI company tax, it shapes what draft legislation looks like next.
On June 10, 2026, Dario Amodei published "Policy on the AI Exponential" on his personal site. It is the most detailed public policy agenda any major AI lab CEO has put forward, and it marks a clear shift for Anthropic: from advocating for transparency measures and voluntary commitments to calling for binding, enforceable rules with government veto power over model releases.
The essay is structured around five areas where Amodei argues governments need to change their approach. This guide covers each one, flags what is verified versus what is still a proposal, and identifies the implications for teams that deploy AI today.
The five pillars of the proposal
The five areas Amodei identifies are:
- Frontier-model safety regulation
- Job displacement and macroeconomic policy
- Accelerating beneficial uses of AI (science, medicine, climate)
- Protecting civil liberties from AI-enabled concentration of power
- Securing democratic leadership in the global AI race
The framing is deliberately large-scale. Amodei argues that AI could produce labor market disruptions that are larger and longer-lasting than any previous wave of automation, and that existing policy frameworks were not designed for that scenario.
Pillar 1: Binding safety testing for frontier models
This is the most significant policy shift in the essay. Amodei is calling for mandatory third-party safety testing of frontier AI models before they can be released, modeled explicitly on how the FAA oversees aircraft certification.
The core proposal: frontier models above a compute threshold must undergo testing by a qualified third party in four domains before they are cleared for release:
- Cybersecurity capabilities (could the model enable serious cyberattacks)
- Biological weapons potential (could it meaningfully assist in creating bioweapons)
- Loss-of-control risks (could the model pursue goals that override human oversight)
- Automated R&D capability that could accelerate the other three
Governments would hold the power to block releases that fail testing. Anthropic committed to putting "substantial financial backing" behind a draft legislative proposal on frontier model testing.
This is a departure from Anthropic's prior public position. Previously, the company focused on voluntary safety commitments and transparency disclosures. Endorsing government veto power over releases is a different posture, and one that puts Anthropic closer to the position of AI safety researchers who have long argued that voluntary measures are insufficient.
For teams using AI: The testing obligation in this proposal targets model developers above the compute threshold, not API users. If something like this passes, you would see the models you access go through a more formal pre-release gate. Releases might become less frequent or slower. Review your reliance on rapid model updates in time-sensitive workflows.
Pillar 2: Job displacement policy and the AI tax proposal
This pillar generated the most press coverage. Amodei argues that AI-driven displacement could be more severe and more persistent than previous technological transitions, and that governments need frameworks before the disruption peaks rather than after.
His proposed toolkit has three components.
Better data collection. Current labor statistics do not track AI-specific displacement. Amodei calls for dedicated tracking of which jobs, sectors, and skills are being displaced by AI, at what rate, and how workers are transitioning. Without that data, policymakers are responding to unemployment figures that blend AI-driven displacement with other causes.
Pro-employment retention incentives. This is the component most directly relevant to businesses using AI. Amodei proposes tax incentives for employers who use AI to augment workers rather than replace them. The details are not finalized in the essay, but the direction is clear: employers who integrate AI without cutting headcount would receive a tax benefit. No mandates, no penalties, just an incentive structure to slow displacement.
UBI as a contingency. If AI-driven displacement becomes large enough to permanently reduce labor demand, Amodei floats universal basic income as one policy mechanism. He suggests it could be financed through taxes on "relevant companies" (the AI companies benefiting from the displacement) or by raising the capital gains tax. He is explicit that UBI is a contingency for a scenario of severe, permanent displacement, not a near-term proposal.
On the same day the essay published, Anthropic announced a $200 million Economic Futures Research Fund to back research into these policy questions and evaluate which approaches are most effective. The company also committed $150 million to a Claude Corps national fellowship placing 1,000 paid fellows in nonprofits to study AI's effects on labor markets. The first cohort of 100 Claude Corps fellows begins in October 2026.
For teams using AI: The retention incentive proposal is worth tracking for two reasons. First, if it passes in any form, companies that already document their AI deployment decisions and staffing patterns will be better positioned to claim those incentives. The documentation groundwork you build for the AI governance checklist is the same documentation that would support an incentive claim. Second, this adds a forward-looking policy context to the current wave of AI employment litigation. The Workday lawsuit and the NYC Local Law 144 audit requirements represent existing legal obligations. The Amodei proposals represent the direction of coming obligations. Both deserve attention.
Pillar 3: Accelerating beneficial uses of AI
The third area is about removing friction rather than adding rules. Amodei argues that AI's potential benefits in biology, medicine, and climate research are large enough that governments should be actively clearing regulatory barriers to research applications, not just managing risks. He points to AI-assisted drug discovery and disease modeling as areas where streamlined regulatory pathways would compound beneficial outcomes faster.
This pillar has fewer specific proposals than the others. The essay calls for dedicated fast-track processes for AI-assisted research applications and argues that the framing of AI regulation should include both risk management and benefit acceleration rather than only the former.
Pillar 4: Civil liberties and concentration of power
The fourth area is a warning about AI-enabled surveillance and information control. Amodei identifies two risk vectors: authoritarian governments using AI to monitor and control populations at scale, and private companies using AI to concentrate control over information in ways that undermine democratic accountability.
He does not name specific countries or platforms, but the implications point clearly at both authoritarian state deployments and large-platform information monopolies. The policy remedies he gestures at are competition law and civil liberties frameworks, though the essay is more diagnostic than prescriptive in this area.
Pillar 5: US democratic leadership in the global AI race
The fifth area frames AI development as a geopolitical competition. Amodei argues that the United States and allied democracies maintaining a technological lead in AI is itself a governance priority, because the alternative is frontier AI development shifting to actors with weaker accountability structures. He endorses US government support for domestic AI infrastructure and coordinated policy with democratic allies.
Why this matters even though it is not law
None of the five pillars are binding on anyone today. So why does this matter for compliance and governance planning?
Three reasons.
Anthropic is funding draft legislation. The essay is not an op-ed. The company explicitly committed to backing a draft legislative proposal on frontier model testing. That means lobbying activity, congressional testimony, and bill text are coming. The positions in the essay are the positions Anthropic's government affairs team will argue for in Washington.
The essay shifts what is politically viable. Mandatory third-party testing with government veto power was, until recently, a position held primarily by AI safety researchers outside industry. An Anthropic endorsement changes the political economy. Watch for versions of it in Senate AI legislation, particularly in national security contexts where bipartisan agreement is more available. The One Big Beautiful Bill preemption debate showed Congress is actively working through what federal AI law looks like; the Amodei proposal fills a gap in that debate.
The labor market policy gap is real and filling quickly. State legislatures have already moved into the space the essay identifies as underserved by federal policy. NYC Local Law 144 requires annual bias audits for AI hiring tools. The Workday litigation is testing employer liability for AI screening decisions. The FTC enforcement actions show federal regulators acting on AI deception even without specific AI legislation. Federal employment policy around AI is a matter of when, not if.
What to watch in the next six months
Draft legislative text on frontier model testing. Anthropic said they would produce one. Senate Commerce Committee and Senate Armed Services are the most likely venues for early hearings.
Economic Futures Research Fund initial findings. The $200 million fund will publish research that directly shapes the labor-displacement policy debate. Initial data in late 2026 could shift the Overton window on retention incentives and influence state-level legislation in 2027.
State-level retention incentive pilots. California, Washington, and New York have large tech workforces and active AI legislation tracks. Any of them could introduce early versions of retention incentives before federal action materializes.
EU convergence signals. The EU is already building a liability framework through the EU Product Liability Directive and Article 50 of the EU AI Act. Amodei's binding regulation proposals are structurally closer to the EU model than to the US historical approach. Regulatory convergence between the US and EU on frontier model rules would be the most significant change to the global compliance picture since the GDPR.
Related Reading
- One Big Beautiful Bill AI preemption: Senate voted 99-1 against it
- FTC AI enforcement actions 2026: all cases analyzed
- Workday AI lawsuit and HR screening checklist 2026
- NYC Local Law 144 AI bias audit employer guide 2026
- EU Product Liability Directive and AI: what deployers must know before December 2026
- AI governance checklist 2026
