In Stockholm last November, a developer asked me a question I couldn't answer cleanly. He was building an autonomous agent that rebalanced liquidity on a Layer 2, and he wanted to know which regulator would come for him first. I gave him the honest version: whichever administration happens to be in office the day the agent breaks something. That answer sat badly with me for weeks, because it described a world where the rules are not a property of the code β they are weather.
Barack Obama said something adjacent on September 14, roughly seven weeks before the 2024 US election. He urged Democrats to move AI regulation up the agenda, warning that without urgent action and a clear plan, the technology would bring dangers. Cointelegraph carried it as a short item β no venue, no link, no policy detail. I have learned to sit with thin source material before drawing conclusions. This one deserves a second look, not for what it said, but for what happened to it afterward.
Here is the timeline the item left out. In October 2023, the Biden White House signed Executive Order 14110, the most ambitious federal attempt to structure AI safety review. By September 2024, California's SB 1047 had cleared the state legislature and was waiting in Gavin Newsom's inbox; he vetoed it on September 29. Four months later, in January 2025, EO 14110 was repealed outright. By July 2025, the federal posture had flipped to an AI Action Plan organized around speed and deregulation. The urgent action Obama called for in September 2024 was, within fifteen months, no longer the law. The instrument of urgency was itself the most reversible thing in the room.
I am not writing this to score a political point. I run a crypto education platform, and the people I teach keep asking me to read regulatory tea leaves as if they were permanent. They are not. The US AI regulation fight demonstrates something the blockchain world understood before it had vocabulary for it: a rule is a promise, and promises are only as durable as the coalition that keeps them. The moment the coalition changes, the promise evaporates β sometimes faster than the code it was supposed to govern ever shipped.
That distinction β promise versus property β is where the two governance camps diverge, and where crypto has something genuinely useful to add.
The safety camp, which Obama was speaking into, works from the precautionary principle: technology brings danger, so act before the danger materializes. The acceleration camp, which won the 2025 argument, works from competitive anxiety: regulation slows domestic builders while rivals keep shipping. Both camps argue about jurisdiction β who writes the rule, how fast. Neither argues about enforcement capacity, which is the actual bottleneck. I have sat in rooms where regulators were asked how they would evaluate a frontier model, and the honest answer was that they lack the compute and the technical staff to do it at scale. A statute can pass on a Tuesday. The ability to verify it does not arrive on Wednesday.
This is the gap the policy debate keeps papering over with the word "dangers." The item never once specified which dangers. Capability risk β misuse, loss of control. Application risk β bias, deepfakes, misinformation. Social risk β displacement, concentration of power. These need entirely different instruments. Lumping them together is not an oversight. Vagueness is a feature of political mobilization and a defect of regulatory design. You cannot write a passing grade for a test you refuse to define.
So the crypto-native reader should ask a different question than "will AI be regulated." The question is whether the guarantees people actually care about can be made into properties of a system rather than promises from an authority. That is why I stopped writing price commentary and started teaching verifiability. I learned to stop preaching and start listening on this point β moral clarity is not an answer to a builder asking which regulator is coming.
Trust is no longer a promise; it's a protocol. We didn't arrive at that line to make a slogan. Trustless systems require trusting relationships; we arrived at it because we watched governance commitments get reversed across three market cycles. The pivot wasn't about politics or products β it was noticing that the thing I could not stop caring about was not price, but whether systems could keep their word after the people who wrote them were gone.
When I advised on security reviews for two DeFi protocols in 2023, the hardest part was never the Solidity. It was getting builders to accept that users would not β and should not β take the team's good intentions on faith. A multisig tells you who signed. It does not tell you whether the signer understood what the transaction did. That gap, between who acted and what was actually intended, is exactly where autonomous AI agents will create the next wave of trouble.
An AI agent executing on-chain does not care whether a regulation passed. It cares whether the call it was handed matches the intent of the human on whose behalf it moves funds. That is a verification problem, not a legislative one. Zero-knowledge proofs, attestation layers, hardware-backed execution β none of these substitute for law, and I won't oversell them. But they are the only tools that survive a change of administration. The EO died. A proof that a transaction matched a stated intent does not die with it.
In 2026 I founded an initiative I called Human-Centric Blockchain and ran a summit in Stockholm with developers and ethicists. The manifesto, The Soul of the Code, made one argument the AI policy conversation keeps missing: blockchain's real contribution to an AI-driven economy is not throughput β it is the ability to verify human intent. Not merely to settle a transaction, but to attest that a human authorized one.
If you want to see where this leads, look at the states. With federal action dead, the rules fragment. California tried SB 1047 and lost it to a veto; New York, Texas, and Colorado have each moved their own bills at different speeds with different definitions of a "covered model." For a builder, fifty different definitions of high-risk AI is a worse outcome than one strict standard, because compliance becomes a map you redraw every legislative session instead of an engineering requirement you satisfy once. The industry that suffers most is not the giant lab with a policy office in Washington. It is the three-person team that ships fast and suddenly needs a lawyer in four states.
Watch where the money moves. Every regulatory cycle creates the industry that sells compliance to the regulated. GDPR summoned a data-privacy cottage industry. Europe's AI Act, in force since August 2024, has already seeded model-audit shops, red-team vendors, and evaluation consultants. Had a federal US safety regime survived, the same dynamic would have played out, and the first customers would have been the largest labs β able to absorb the cost and quietly enjoy the barrier it creates. I have seen this movie inside crypto too. The listing standards exchanges adopted voluntarily became a filter favoring projects with legal budgets over projects with better code. Rules rarely arrive neutral.
There is a version of this that matters specifically to a market like this one. In a bear market, the scarce resource is not attention, it is runway β and runway spent reacting to regulatory weather is runway not spent building. The protocols that survive the next eighteen months will be the ones whose core guarantees do not depend on any regulator's mood. When I ran the Yield & Connect meetups in Stockholm during the 2020 DeFi summer, the whole point was that liquidity pools could rebuild community trust that institutions had spent decades eroding. That argument was never about yield. It was about whether a system could make a commitment that outlived the people who made it.
Which brings me to the contrarian point, because the crypto reflex here is wrong and I want to say so plainly. The reflexive industry position is that AI regulation is somebody else's fight β stay quiet, avoid the crossfire, keep our heads down until the cycle turns. That is a mistake, and not for the reason most people assume. It treats all regulation as cost. Good regulation is not a cost to incumbents. It is a moat.
A clean federal AI safety standard, applied uniformly, would hurt small developers β who cannot staff an audit function, cannot fund continuous evaluation, cannot wait eighteen months for a license β far more than it hurts the labs that can. The safety camp's own "dangers" rhetoric, codified, would freeze the frontier exactly where it currently sits. That is the blind spot. Crypto's instinct is to fight every regulation as hostile. The durable fight is subtler: push for rules that are verifiable, transparent, and cheap to satisfy, and write them into protocol commitments rather than administrative discretion the next administration can erase with a signature. A transparency regime you can satisfy from an open repo survives. A licensing regime with discretionary approval does not.
The reversal of EO 14110 is the proof. It was not defeated in court. It was not defeated by argument. It was defeated by an election. Any crypto firm that built its strategy around one administration's regulatory clarity built on sand β and had eighteen months of warning.
None of this means the safety camp is wrong. It means the safety camp is arguing for guarantees it cannot currently enforce, using instruments that expire. I have watched the crypto industry make the same mistake from the other direction β begging for a favorable rule instead of building a system that needs no permission. The AI world is about to learn what we learned the slow way.
I keep coming back to the developer in Stockholm and his autonomous agent. Somewhere in the next two years, an AI system will move real capital on-chain without a human clicking confirm, and something will go wrong. When it does, the first question will not be which statute applies. It will be whether anyone can prove what the agent was authorized to do. That is the question our industry is uniquely positioned to answer, and the AI regulation debate is not even asking it. We should be building the answer while the argument is still about jurisdiction. Because the argument will end. The proof, or the absence of it, will remain.
