From the chaos of 2017, we forged a compass — and I have spent the years since watching, with quiet dread, as people walk back into the woods. This past cycle, Dario Amodei, the chief executive of Anthropic, outlined a three-step strategy for the responsible global development of artificial intelligence, calling for international cooperation and safety alignment as a bulwark against a technology that no single laboratory, and no single nation, can govern alone. The announcement landed with the polished inevitability of a keynote, and the coverage reduced it, almost instantly, to a single sentence: a powerful man proposed that the world should be careful. What unsettled me was not the proposal. It was how readily we accepted that "careful" would be defined by the same institutions racing to build the thing we are all meant to be careful about. I have audited enough systems to recognize the shape of this move.
Anthropic is not a fringe voice, and pretending otherwise would be its own dishonesty. It ranks among the most highly valued AI laboratories in the world; its Claude models sit in the same tier as OpenAI's GPT line and Google's Gemini. When its chief executive speaks about safety, the market listens — not because he is disinterested, but precisely because he is positioned. That is the context the breathless summaries omit. A governance framework proposed by a competitor is simultaneously a public good and a strategic asset, and the two cannot be cleanly prised apart.
The regulatory terrain has shifted beneath everyone's feet. The European Union's AI Act entered into force in 2024. The United States issued its executive order the same year. China implemented its own interim measures for generative AI services. Each jurisdiction is raising a cathedral of rules, and none of them can see over the others' walls. Amodei's invocation of global cooperation answers that fragmentation directly — and fragmentation, as I have argued for years about liquidity, is often less a natural condition than a narrative somebody finds it profitable to sustain.
Anthropic, notably, already owns a template. Its Responsible Scaling Policy sorts models into safety levels, each carrying escalating evaluation requirements before deployment. To a trained eye, a three-step strategy reads like an attempt to lift that internal scaffold into an external standard. That is a genuine ambition. It is also a bid for the authority to decide where the ladder's rungs are placed, and who must climb them.
Here is the heart of the matter, and it deserves to be said plainly rather than cushioned: the power to define "safe" is the power to define the market. Once an industry accepts a shared threshold, every product must clear it or perish. The threshold becomes a moat, and its author becomes the gatekeeper. I do not say this to accuse. I say it because I watched the same mechanism run its course in a smaller arena, and the ending was not what the founders promised.
In 2020, during the first DeFi Summer, I stopped trading and started verifying. I manually checked more than two hundred protocols against open-source standards and built a simple, human-readable dashboard so that non-technical users could see what they were actually trusting. The community grew to ten thousand members, and their incident rate fell by eighty percent. What that experience taught me was not that audits are magic. It was that accessibility, not cryptography, is the true frontier of decentralization — and that a standard is only as trustworthy as the number of independent people who can verify it. A safety framework designed by one lab and validated by no one is not a standard. It is a sentiment with a logo.
Two years ago I stood in front of a London financial forum and challenged a room of institutional investors on the centralization hiding inside their custodial solutions. True ownership, I argued, is non-negotiable, and a self-custody education module followed from the one firm willing to listen. I tell this story because the same custodial logic is now being proposed for AI. A handful of labs — holding the models, the weights, and the safety standards — would become the custodians of intelligence itself, asking us to trust their stewardship for the same reason banks once asked us to trust theirs: because it is convenient, and because the alternative sounds frightening.
This is where the AI governance conversation and the crypto conversation I have lived inside for fourteen years finally collide, and where most readers have not looked. The AI industry is proposing to answer a coordination problem with a centralized declaration. We already know how that pattern resolves, because we have run it on our own chain. From 2022 onward, I published a thesis arguing that sustainable ecosystems require emotional and social capital, not merely economic incentives; three major DAOs cited it in their charter revisions. Many of the DAOs that ignored that lesson did not fail because their code was broken. They failed because governance was captured by exactly the parties who promised to protect it.
So when I hear that a three-step strategy will "ensure" global safety, my cryptographic instinct does not ask what the steps are. It asks who holds the keys to the verification. And here the omission is glaring: there is no mention of an independent, adversarial, cross-border mechanism of attestation — nothing resembling the zero-knowledge proofs we now use to verify computation without trusting the machine that performed it. The infrastructure for verifiable AI is arriving faster than most policymakers realize. Systems are already proving, in production, that a model ran as claimed, that training data met a declared standard, that an inference respected a stated constraint. We do not need to trust a lab's safety report; we can cryptographically verify it. The technology exists. The strategy, as described, does not use it.
Last year I launched the Human-Centric AI Ledger, a cryptographic protocol for verifying the origin of AI decisions, and it drew two million dollars in grants from ethical technology funds. We are not proving that a model is good. We are proving that a model is what it claims to be — that the decision you received came from the process described, not from a black box wearing a safety badge. That distinction is the entire argument. Governance framed as a promise asks you to believe. Governance framed as a proof asks you only to check.
There is a hard limit worth naming, and it is not fashionable. The data availability layers we spent the past two years celebrating are already straining under real load; the blob space that post-Dencun rollups treated as effectively infinite will be saturated within roughly two years, and when it is, gas fees will climb again — the quiet bill for optimism nobody priced. If AI governance is to be anchored on-chain through attestations and provenance proofs, it will inherit every one of these constraints. Verification is not free, and anyone who tells you the cryptographic layer scales linearly with ambition has never watched a network's cheapest block fill up.
Nor should we repeat our own mistakes about which layer carries which burden. We tried, for a season, to load governance and culture onto Bitcoin through inscriptions and Runes — hauling freight in a Rolls-Royce — and we learned, at some cost, that a machine built for one purpose rarely thanks you for being repurposed. AI governance placed naively on the wrong substrate will produce the same disappointment: impressive throughput, negligible sense. The question is never whether it can be put on-chain. The question is whether the chain in question is structurally suited to hold it.
And the geopolitics cannot be wished away. China has advanced its own global AI governance initiative; the United States tightens export controls quarter by quarter. A framework authored in San Francisco will be read in Beijing as an instrument of soft power long before it is read as a sincere plea for coordination. The realistic outcome is not one global standard but two — a bifurcated regime in which models and markets drift apart, each carrying its own definition of responsibility, neither speaking to the other. Any honest three-step plan must contain an answer for that, and the summaries we were handed do not.
Let me test my own skepticism, because a moral-first reading demands it. Perhaps Amodei's three steps are exactly what a fractured world needs — a credible, well-resourced actor naming the danger before it arrives, using a competitor's platform to force rivals toward their own safety commitments. That reading is generous, and it may even be true. The pressure he places on other labs is real, and a race toward safety is preferable to a race without one. But pragmatism demands we separate the useful signal from the flattering story. A framework that is authored, hosted, and interpreted by its author is a pledge, not a protocol. Trust is not a metric; it is a memory we share — and no laboratory, however sincere, can manufacture that memory for seven billion strangers by writing it into a terms-of-service document.
Which leaves the uncomfortable contrarian conclusion. The three-step strategy will likely succeed at precisely the goal the market rewards — positioning Anthropic as the responsible leader of a governance era — while failing at the goal the world actually needs, which is an enforceable, independently verifiable, geopolitically durable standard. Those two outcomes are not opposites. They can coexist, and history suggests they usually do. The failure will not look like a collapse. It will look like a summit, a communiqué, and a quiet set of walls raised just high enough that the well-capitalized clear them and everyone else does not.
I want to end somewhere better than that, because cynicism is the cheapest form of security. Amodei is not wrong about the danger, and the engineers building decentralized verification are not naïve about the limits. Somewhere between a keynote and a proof lies the real work: making safety measurable by anyone, anywhere, without a lab's permission. That is not a utopia. It is an engineering discipline we already practice — in auditing contracts, in verifying rollups, in proving state without revealing it. If the AI industry is ready to be governed, it must be ready to be checked by parties it did not appoint. The step missing from every three-part plan is the one that lets a stranger on the other side of the world confirm you told the truth. When a strategy finally includes that step, I will believe it. Until then, I keep my compass.