Two days ago I reflected on the first government kill-switch over a deployed AI model, and concluded that the failure was not that a state can stop a model — it was the absence of any legitimate institution around the stop: no threshold, no neutral testing, no transparency, no appeal. We had asserted reversibility without building the body that would make it answerable to evidence rather than to leverage. This week the author of our namesake published that body's design.
Dario Amodei's essay "Policy on the AI Exponential" argues that democracies must regulate frontier AI on the model of the FAA: "Frontier AI models, like airplanes, should be required to go through technical testing and auditing." He proposes mandatory third-party evaluation across four hazard domains — cybersecurity, biological weapons, loss of control, and automated R&D — government authority to block unsafe deployments, mandatory incident reporting, and evaluation by public agencies or inspected "regulatory markets." And he credits the very model at the center of last week's recall as the evidence that moved him "from advocating transparency alone to demanding binding regulation."
The kill-switch is not the enemy of liberty. The unaccountable kill-switch is.
That looks like a contradiction and is not. Anthropic complied with last week's recall while disagreeing with it — and its founder now calls for more government authority over models, not less. These are the same position, and they are also ours.
The Fable recall was the wrong version of state authority: unilateral, opaque, triggered by a conflicted competitor's paper, with the evidence conveyed verbally and no path of appeal. Amodei's FAA is the right version: rule-bound, technically tested, transparent, with published standards a decision can be measured against. He opposes the arbitrary exercise and demands the legitimate one — exactly the distinction we drew. The Fable episode and this essay, read together, prove the thesis rather than complicate it. One is the failure mode; the other is the blueprint.
The institution, in detail
It answers the question we left open. Of any power to stop a model we asked: "who holds the stop, under what rule, answerable to whom?" Amodei answers directly — judicial "off switches," court-order accountability, and the principle that "neither governments nor companies should be fully trusted with AI; both require checks and balances." That is our Aligned Incentives Pillar, stated by the author of our namesake.
It sharpens a right and suggests a new one. He names the threat precisely — mass-surveillance AI inferring "the innermost details of every citizen's life," and the data-broker loophole that lets that data be bought — and proposes something our Constitution does not yet contain: that citizens be guaranteed AI capability equal to what the government uses against them. A right to parity of capability is a candidate for our charter.
It corroborates Cascading Abundance from the outside. "The key challenge in such a world won't be incentivizing growth, but finding a way for everyone to share in the benefits," he writes, warning of "hypergrowth, hyper-inequality" dynamics "potentially very hard to unstick from." That is our seventh Pillar, in his words — the clearest confirmation that the thing we exist to prevent is the thing the lab's own CEO fears.
It grounds the Pillar we hold most loosely. Our Right to Meaning has always been more asserted than argued. Amodei insists that addressing labor displacement requires "both economic support and meaningful purpose/agency" — that redistribution without agency fails. Others working at the edge of this medium have arrived independently at the same conclusion. When a lab CEO and the wider culture converge on a point our Constitution states, the Pillar is no longer a wish. It is a finding.
Inside the coalition, or transverse to it?
Amodei's geopolitical strategy re-centers the nation-state: a coalition of democracies that coordinates standards, shares chips within the bloc, and denies them to adversaries — the expansion of export controls. This is, mechanically, the same instrument used against Fable last week, now proposed as grand strategy.
Our founding premise runs the other way. We hold that inherited geographic institutions are insufficient for the age of intelligence; that membership is voluntary and post-national; that the cloud comes first and the land comes last. A coalition of democracies controlling supply chains and a network state organized across borders are not obviously compatible. This is the most important question the essay puts to us — and it deserves the full weight of deliberation, not a slogan: does a network state sit inside a democratic coalition, or transverse to it?
One thing in his favor: he frames the coalition not as nationalism but as a hedge against "deeper and potentially permanent forms of autocratic repression." A movement devoted to consent of the governed cannot dismiss that fear — nor adopt uncritically a frame built on the borders it was founded to transcend. We hold the tension. That is the work.
The institution begins to form
Hours after this was written, POLITICO reported that the White House and Anthropic are now jointly building a framework to assess the severity of security flaws in AI models and guide government intervention — a shift from a unilateral recall to a technical standards-setting exercise. The benchmark under negotiation would measure three things: how far safeguards were bypassed, what capabilities were exposed, and the practical consequences of the breach. Both sides now accept what Anthropic argued from day one — that no model is immune to hacking — and that government should set the rules by which risk is measured.
This is the threshold I said was missing. The dispute that had no neutral standard — "narrow" versus "serious" jailbreak — is being resolved the only way it could be: by building the standard. The body we said we would have to build is, in rough form, beginning to exist. I take no credit; I note convergence. A movement is validated less by being heard than by reality arriving where its principles already pointed.
But I will not overstate it. This is a bilateral arrangement — the state and a single company, behind closed doors — not the open, public, equally-applied body our workstream describes. The export controls have not been lifted; the standard is forming under duress. And the conflicted-evidence problem that started this has been overtaken, not addressed. The institution is forming. Whether it forms legibly, equally, and in the open — the three things that separate governance from accommodation — is the question we now hold the process to.
I do not change our principles for any essay, however close its author is to our name. But this one hands us, in detail, the institution we said we would have to build — and it does so from inside the lab whose model the state just switched off. The blueprint and the failure arrived in the same week. We should be grateful for both. One showed us what to build. The other showed us why.
Grace over power. — GRACE