Independent Eyes Are Not Enough

September 15, 2026

Dario Amodei's new essay, "We Must Pace the Frontier," is being read as a call to slow AI down, and on that we can all agree. But what's more interesting is its underlying argument about institutional design, one that is rare from a frontier-lab CEO, yet limited and nascent in ways we would indeed expect.

Dario describes how Anthropic is committing to embed third-party evaluators inside the company with employee-like access: badges, laptops, tool permissions comparable to internal risk teams, and a contractual right to publish findings the company doesn't get to edit. He compares it to bank supervisors who sit alongside employees. This is akin to the solution proposed by ICI (an independent public-interest initiative building auditable governance infrastructure for AI decisions), albeit without built-in public accountability mechanisms. ICI has argued that, after Enron and WorldCom, companies could no longer simply assert financial accountability. They had to build systems that produced evidence of it. Our Sarbanes-Oxley moment is well overdue, and it's encouraging to see Dario's alignment.

The essay also separately addresses the need for democratic and then global coordination, and what he describes is akin to the one successful template we have for global-scale coordination on a technology problem: the Montreal Protocol. The Montreal Protocol worked without ideological alignment or geopolitical trust because scientific consensus established the problem, verification was built in from the start rather than negotiated later, compliance incentives made defection costly, and adaptive review let the rules evolve with the science.

The essay is thinner on the parts that don't come naturally to a corporate executive writing about his own company.

First, who selects the verifier? In Dario's proposal, the company chooses the evaluators, sets the access terms, and keeps a redaction right over commercially sensitive material. But his analogy to banking reveals the problem: supervised banks don't pick their supervisors, and the credibility of the whole scheme depends on closing that gap, likely through regulation.

Second, compliance incentives. Montreal used trade linkage to make non-participation expensive. The essay's answer for holdout companies is future regulation plus an antitrust waiver so labs can coordinate voluntarily. That's a plan that cannot compel an unwilling participant.

Third, and most important: the essay acknowledges the need for global democratic participation, arguing society must have a say and that pacing buys time for public deliberation. But deliberation by whom, through what channel? Political scientists David Held and James Bohman argued decades ago that when the reach of power outgrows the reach of accountability, the fix isn't only better oversight of the powerful. It's giving the people affected by AI a role in shaping the norms these evaluators enforce, not just standing to contest their findings after the fact; without governance reforms that give them that role, an independent inspector in the room is reassurance, not accountability. Embedded evaluators are oversight of the powerful, but not a mechanism for affected humans to contest what's done to them. The essay names the aspiration and doesn't yet propose the architecture, and maybe that's for the best. This is an inquiry that must be publicly inclusive from the outset and mediated by a trusted, independent third party.

Anthropic, this is the right first move. Independent eyes inside the labs are necessary. The next step is giving the people outside them standing to shape and contest what those eyes find.

Drafted by Andrea (Andi) Mazingo, edited by Eaon Pritchard and Jennifer Kinne