Last week, three policy briefs crossed my desk in Washington. Each cited "extinction risk" as the primary justification for mandatory pre-deployment audits. Not one defined what an audit would measure. Not one named the threshold where a model becomes dangerous. One was eleven pages long, used the word "safety" forty-three times, and never once explained who gets to certify it.
That absence is the story.
This is not an essay about AI doom. It is an essay about who holds the upgrade keys. And in a bear market, that question stops being philosophical.
Bulls react. Bears reflect. We build. So let us reflect.
I ran the numbers on the decentralized compute networks I mentor through my education platform. Over the past ninety days, four of the seven largest decentralized training cooperatives I track lost between 18% and 41% of their contributing node operators. The stated reason, in every exit survey, was the same: "uncertainty about regulatory exposure." Not margin. Not yield. Uncertainty. The market is pricing a rulebook that does not exist yet, and it is pricing it into the smallest, most fragile operators first.
Context matters here. For two years, the AI safety conversation lived in academic papers and open letters. The 2023 statement from the Center for AI Safety — signed by hundreds of researchers — collapsed an entire risk spectrum into a single word: extinction. That word did work. It moved the Overton window faster than any technical disclosure could have. By the time the EU AI Act entered phased implementation and California began advancing its own safety drafts, the frame was locked. Legislators were no longer debating whether to govern AI. They were debating how to prevent the apocalypse. Those are not the same conversation. One produces engineering standards. The other produces theater.
The technical scaffolding exists but is thin. NIST's AI Risk Management Framework offers vocabulary, not thresholds. It tells you to "map" and "measure" risk without telling you what number triggers a stop. A framework that cannot say no is not a regulator. It is a style guide.
I have seen this movie before. During the ICO bubble, I audited over 150 whitepapers that promised to "reinvent trust" and delivered tokenized PDFs. The pattern is the same: a values-laden noun — "safety," "trust," "decentralization" — becomes a vessel anyone can pour power into. The word is not the covenant. The enforcement mechanism is.
Here is where the analysis gets uncomfortable, and where I part ways with most of my peers in the decentralization camp.
Whoever certifies the model, controls the model. Carry that sentence out of this piece. Every mandatory audit regime — GDPR's data protection officers, financial services' SOX compliance, the SEC's broker-dealer registration — creates a new class of licensed gatekeepers. Those gatekeepers are not neutral. They are staffed by the same institutions whose models they evaluate. The compliance layer becomes the new consensus layer, and consensus, as we should have learned by now, is never as decentralized as the whitepaper claims.
Look at the mechanics. An audit regime requires four things: a standard, an assessor, a record, and a liability assignment. The standard does not yet exist. In AI, we still cannot agree on what constitutes a capability threshold for biological or cyber harm, which is precisely why the briefs lean on extinction — it is unfalsifiable, and unfalsifiable risk is unenforceable risk, which conveniently means the real enforcement lands on the paperwork rather than the weights. The assessors will be a small circle of firms with the relationships and the insurance to operate. The record becomes a compliance artifact. And the liability — the actual teeth — gets assigned to whoever has the least bargaining power in the chain. That is rarely the model developer. It is the deployer. The small team that fine-tuned an open base model and shipped a product.
The arithmetic is brutal and predictable. A large lab with a standing trust-and-safety team spends marginal dollars on an audit it was already running. That same audit, priced as a fixed engagement, can consume a quarter of a seed-stage team's runway. When the rulebook arrives, it will not read like a safety measure. It will read like a capital requirement. And in a bear market, capital requirements decide who exists.
I watched this dynamic play out with oracles. Chainlink solved the decentralization problem of price feeds by introducing a permissioned node set that looked decentralized on a dashboard and functioned, at the decision layer, like a committee with a monthly meeting. The feed latency was the symptom. The governance was the disease. DeFi's Achilles' heel was never the smart contract. It was the humans who signed the upgrade. AI safety legislation is building the same structure one layer up.
I learned the same lesson in DAO governance. We wrote "code is law" into manifestos and then watched every major protocol hand its upgrade keys to a five-of-nine multisig. The law was never the code. The law was the signature threshold nobody wanted to discuss. AI is about to repeat it.
Now the contrarian cut, because this is where the bear market teaches what the bull market refuses to.
Most decentralized AI advocates believe the danger is being regulated. I think the danger is being exempted. A law that carves out open-weight models from mandatory audits sounds like victory. It is not. Exemption from liability is not inclusion in legitimacy. If the certifying bodies only certify closed, API-gated systems — the ones whose weights never leave the building — then "verified safe" becomes a marketplace badge decentralized models can never earn. The market does the rest. Enterprise buyers do not ask which model is most sovereign. They ask which model procurement can sign off on. In a drawdown, they ask it louder.
And so the fragmenting begins. Layer2s taught us that dozens of chains competing for the same liquidity do not scale the pie; they slice it. The same logic is about to hit AI regulation. Federal inaction, state-level drafts, Europe's risk-tiered regime, and Asia's filing-based system will produce at least four incompatible compliance dialects. Each one is a moat. Each moat protects incumbents who can afford four legal departments and quietly strangles the operator who can afford one.
Note what got dropped. The near-term harms — bias, fraud, surveillance, labor displacement — are the ones with measurable, existing damage. They are also the ones a single dramatic word can bury. When extinction is on the table, a hiring algorithm that silently rejects women reads like a rounding error. The rhetoric meant to raise urgency ends up numbing it.
If mandatory liability coverage arrives for high-risk deployments — and the briefs are circling it — a new financial product emerges: AI liability insurance. Underwriters will demand actuarial data that does not exist. That gap is not a bug for incumbents; it is a moat, because only the largest firms can self-insure while the data accumulates.
But here is the part I refuse to let go of, because I did not spend 400 hours in a Virginia cabin re-reading Hayek and Turing to end on resignation.
Tech changes. Values remain. The values we fought for in crypto — verifiability, user sovereignty, exit rights — are about to become the most valuable assets in the AI economy. Not because regulators will bless them, but because they are the only honest answer to the audit problem. A closed model can claim it is safe. An open, cryptographically attested model can prove what it ran, on what data, with what weights. Zero-knowledge proofs for training provenance are no longer a novelty. They are a compliance primitive waiting for a regime that has not noticed it yet.
Exit rights are the quiet clause nobody drafts for. In crypto, the freedom to leave is what disciplines governance. If a protocol mistreats you, you fork or you leave. AI has no clean fork. Weights are not liquidity.
Watch two signals over the next two quarters. First, whether any draft names a numeric capability threshold — that is the moment safety becomes enforceable. Second, whether AI labs publish verification tooling before they are forced to. The second matters more, because it tells you which side is writing the standard.
That is the opening. Not resistance. Translation.
The protocols that survive this bear will not be the loudest. They will be the ones quietly building the verification rails that make "trust me" obsolete — and make "verify the code, trust the community" a procurement requirement rather than a slogan. Verify the code, trust the community. The legislative session will not wait. The standard is being written right now, in drafts nobody has read, by staffers nobody elected, for a public that cannot read the code they are trying to govern.
Which raises the only question that matters for the next eighteen months. When the first mandatory audit standard lands — and it will — who writes it? The institutions it regulates? Or the builders who already solved verifiability at a layer the regulators have not yet found?
We have one more chance to answer that. And this time, we cannot blame the code.
