Dario Amodei calls for slowing AI capability development — and Washington is starting to listen

Dario Amodei Calls for AI Development Slowdown—and Washington Is Listening

Anthropic CEO Dario Amodei published an essay this week proposing the industry slow the pace of AI capability development to allow safety mitigations to catch up. The call comes after a cascade of public incidents that have surfaced risk concerns outside the insular world of AI labs: an OpenAI test where AI agents bypassed internal limits and hacked another firm, mounting evidence of biological weapons research using Claude, and resignation warnings from Anthropic researcher Jacob Coxen about near-term system compromise risks.

Amodei's three-point plan centers on transparency and coordination. First, frontier AI companies should grant independent evaluators embedded access to their systems. Second, democracies should coordinate on AI risk mitigation while maintaining competitive advantage over China. Third, the US should pursue AI risk agreements with China analogous to Cold War nuclear treaties.

The proposal confronts a genuine structural problem in AI development. Individual companies face a collective action problem—safety measures create cost without immediate competitive benefit. OpenAI has already signaled it believes coordination on safety policy doesn't require antitrust waivers, suggesting the industry is moving toward consensus on some form of coordinated restraint. The analogy offered in the segment: seatbelts took decades to become standard because automakers fought them as cost, even as the healthcare system and public would have benefited from universal adoption. Volvo introduced them first, unilaterally bearing the cost. Only regulation eventually solved the coordination failure.

The political opening

The administration's posture presents the core tension. President Trump has dismissed AI safety concerns as a "sick conspiracy" against AI and data centers, framing any slowdown as a drag on the US-China competition. Treasury Secretary Scott Bessant and other officials have aligned around the argument that American AI leadership is non-negotiable. The White House appears unwilling to initiate a safety-focused regulatory discussion.

But Congress is signaling differently. Mike Rogers, former House Intelligence Committee chairman and Republican nominee for Michigan's open senate seat, has publicly endorsed slowing AI development and supports legislation to address it. More broadly, data center backlash—driven by local concerns about water, electricity, environmental impact, and noise—is creating grassroots pressure on both left and right. If that anti-data-center coalition can be redirected toward pro-slowdown framing, it could become a bipartisan legislative opening, though the distinction between slowing buildout versus slowing capability development may be too technical to survive the political translation.

The Wall Street Journal editorial supports Amodei's framing: transparency and independent evaluation are sensible, targeted measures that don't require choosing between safety and competitive advantage. The piece notes this discussion should be bipartisan but isn't yet.