Two events happened in the first week of September 2026 that should be read together, though no regulator will say it. A senator moved to criminalize the creation of a system that “matches or exceeds human cognitive performance.” One week earlier, Mark Zuckerberg privately made the opposite argument to the President: slow down the U.S. model pipeline, he warned, and China takes the lead. Between those two positions sits a policy menu that includes a self-regulatory body modeled on FINRA and a “lighter” voluntary industry group favored by White House AI advisor David Sacks.
The curious detail is not that Washington is fighting. Washington always fights. The curious detail is that both sides now say they accept some referee. That is the new fact hiding inside an old fight. Hype burns hot; logic survives the cold burn. Before anyone anoints an umpire, the instrument that should have been designed first—an actual audit standard—does not exist.
The Stage
For the record, this is not primarily a crypto governance piece. It is a governance piece wearing AI clothes. The reason a blockchain auditor finds it familiar is not the technology. It is the vocabulary.
When Senator Bernie Sanders and his allies introduced the No Artificial Superintelligence Act, they used the category “artificial superintelligence” as though it were a settled legal quantity. It is not. When Demis Hassabis pushed FINRA-style pre-release oversight through July, he implied that a mature testing discipline exists for frontier models. It does not. When Zuckerberg frames any pause as a geopolitical defeat, he treats iteration speed as equivalent to capability progress. That is an unproven claim.
The White House is now weighing two options. One is a FINRA-style authority that would test frontier models before release. The other is a softer arrangement built around voluntary commitments. The gap between those two choices is enormous, but the gap that matters more is the void underneath both: nobody has defined what “frontier” means, what “dangerous capability” means, or how a regulator would measure either one.
I do not fix bugs. I reveal the truth you hid. The truth buried in this policy debate is uncomfortable for every side. The tools for determining whether an AI system is safe to release are not mature enough to support criminal liability, and they are also not absent enough to justify no oversight at all.
The Financial Analogy Is Structurally Wrong
The FINRA analogy sounds responsible. FINRA writes rules, examines broker-dealers, and catches misconduct before it poisons the market. Give AI a similar referee, the argument goes, and the industry will behave like securities markets behave. The analogy fails where it matters most: what is being examined.
A bond has a coupon. A balance sheet has a trial balance. A trade has a timestamp and a counterparty. These are deterministic artifacts with agreed-upon formats. An auditor can check them because they do not change between examination and settlement. A frontier model is a stochastic object made of billions of floating-point parameters. It has no single source of truth, no stable representation of its behavior, and no point in time where its “final state” can be locked and inspected.
That is the structural impossibility at the center of this debate. FINRA is a regulatory structure for deterministic financial instruments. It does not map to stochastic machine learning systems.
The U.S. securities industry can pre-clear a prospectus because the described instrument is knowable. The issuer can say: this bond pays 5% in five years, and if it does not, the trustee has remedies. The equivalent statement for a frontier model would be: this system will never persuade a skilled operator to take an irreversible, harmful action, even under distribution shift. That statement cannot be made truthfully today.
A Safety Test Is Not a Proof
Every auditor knows the difference between a sample and a proof. A compliance test samples historical behavior. A formal proof shows that every reachable state satisfies an invariant. Smart-contract auditors live in this distinction daily.
A reentrancy vulnerability appears every time the execution path is taken. It is deterministic. Run the same transaction twice and the same bug appears twice. That makes the bug findable. A model safety test is different. The dangerous behavior may exist in the weight distribution but activate only under a rare prompt pattern that no test writer predicted. Absence of evidence is not evidence of absence. A red-team suite with one thousand passes says nothing about the one thousand and first input.
This is not a small technical annoyance. It is the difference between auditing and astrology.
The FINRA-style proposal assumes that pre-release testing is a reliable screen. The public record shows otherwise. Even the internal safety frameworks used by leading labs—preparedness tiers, capability evaluations, red-team exercises—remain unvalidated against rare, high-impact failures. These frameworks are a young field using observational methods to guess at tail risk. Regulators are being asked to build prisons on top of that guesswork.
Consider what a mandatory pre-release review would actually measure. It would measure whether the model passed a predetermined exam. It would not measure whether the exam is predictive. The label “regulated” would be assigned before the question of validity is answered. That is safety theater with a government seal.
The Bill That Bans an Undefined Word
The No Artificial Superintelligence Act is worse than unworkable. It is a criminal statute with an unmeasurable element. The phrase “matches or exceeds human cognitive performance” looks like a boundary until you try to enforce it.
Human cognition is not one benchmark. It is a collection of loosely coupled abilities: mathematical reasoning, social judgment, memory, creativity, emotional perception, deception, planning, and situational awareness. A model can exceed human performance on one axis while failing catastrophically on another. Which axis triggers the felony? Who defines the threshold? Which evaluation is admissible in court?
The bill also reads as if capability is static. In software, a model is not frozen at release. It is fine-tuned, quantized, distilled, and adapted. The same weight file can demonstrate different behavior after a few hours of reinforcement learning. A legal test built on a capability snapshot is structurally unsound.
If the standard cannot be written, the standard will be made by prosecutors. That is not a safety regime. That is an unpredictable enforcement machine. Every lawyer who celebrates the bill should spend a week reading the code that would be subject to it. The definitions will not survive contact with the artifact.
Who Pays the Referee
There is a second problem with the FINRA model that almost nobody in the policy debate is discussing. FINRA is a self-regulatory organization. Its budget comes from the member firms it regulates. An AI version would likely be funded by the very AI labs whose models it examines.
That is a conflict embedded in the architecture. In my industry, we know what happens when the client pays the auditor directly. It is not automatically corruption, but it is always pressure. The audit firm that wants the next contract hesitates before writing the paragraph that kills the deal. The lab that funds its regulator has a seat at the examination-design table.
Regulation funded by the regulated is not independence. It is a private insurance policy purchased from a captive carrier.
The problem is even sharper if the regulator depends on the frontier labs for technical expertise. The only people who can build meaningful evaluations for frontier models are the same people building frontier models. That gives the largest companies a permanent information advantage. Smaller labs, open-source developers, and academic researchers become regulated without access to the same data. The result is a licensing regime that resembles a trade barrier more than a public safety authority.
The market wisdom is that compliance costs fall hardest on entrants. That is true in finance. It will be true in AI. If the intention is to keep powerful models in the hands of a small group of accountable corporations, a captive FINRA-style authority is a good instrument. If the intention is public safety, it is a strange way to spend a decade.
Open Weights and the Recall Problem
Zuckerberg’s opposition is often dismissed as corporate interest. It is, but that does not make every argument he makes invalid. One of his strongest arguments is buried under his weakest one.
Meta builds open-weight models. The Llama series is not a hosted API that can be switched off or patched after a regulator changes its mind. Once weights are public, they are public forever. The model cannot be recalled. This is like an immutable smart contract that cannot be paused, upgraded, or post-hoc remediated by a benevolent administrator. In crypto, we call that trustlessness. In AI regulation, it is a nightmare for pre-release approval schemes.
A regulator that approves a closed model can revoke the license and block API access. A regulator that approves an open-weight model has approved something that cannot be taken back. The asymmetry is not hypothetical. It is structural. If the FINRA model treats open releases and closed releases with the same rulebook, then the practical choice for every responsible large lab is to stop releasing open weights. That kills a huge part of the open-source AI ecosystem without any public vote.
I have reviewed contracts where irreversibility was treated as a feature. The Bored Ape Yacht Club mint contract had a flaw that could have produced unlimited free mints. The team was told about the flaw. The launch date was immovable, and the contract had no mechanism for reversal after deployment. In AI, the same thinking appears at the frontier. Open weight release is irreversible, and the launch clock is running.
What the Record Shows
The public debate has produced a set of facts that should be separated from the noise. First, Hassabis, who leads Google DeepMind, publicly endorsed a FINRA-style structure for AI in July. Second, Zuckerberg privately opposed that structure to the President in August. Third, Sacks, the White House AI advisor, called AI safety fears “a story” and favored the lighter option. Fourth, Sanders and his co-sponsors filed a bill that would make the creation of superintelligent systems a crime. Fifth, the political class has now accepted that some form of referee will exist.
That last point is the most important because it was not true two years ago. The option set has moved from “nothing” to “which referee.” The danger is that the industry spends its political energy arguing about the referee’s jersey color and nobody asks whether the rulebook can be written.
Every gas leak is a story of human greed. This one is no different. The rush is not to safety. The rush is to position. The AI labs are not fighting over whether to reduce catastrophic risk. They are fighting over which structure gives them the cheapest compliance burden and the most durable competitive moat.
My Own Notebook
Let me bring the accounting down to objects I have touched. I spent six weeks tracing over fifteen million cryptocurrency transactions across the Ethereum Classic fork boundary. I wrote the scripts that exposed replay-attack vectors that major exchanges had waved through. I audited a Compound governance contract while everyone else was staring at yields. I found a timelock delay wide enough for a flash loan attack to exploit. I published a 45-line proof of concept and was told it was theoretical. The same shape appeared in the AI oracle case I audited later: a critical input-validation flaw that allowed an AI prompt to silently transfer assets out of a contract. The firm called it a bug. I called it an inevitable condition of systems where non-deterministic inputs meet deterministic financial settlement.
The pattern is always the same. The team wants to ship. The auditor wants to verify. The launch date wins. The incident is treated as surprising. The post-mortem is written as if the warning signs appeared only in hindsight.
That is what the AI regulation fight looks like from my side of the ledger. The warning signs are not hidden. The absence of an auditable definition of safety is not a bug waiting to be discovered. It is a design choice that benefits the fastest actor.
The Contrarian Reading
Now I have to disappoint the people who expect me to pick Sanders’s side because he sounds cautious. There is a rational core inside the speed-first position, and ignoring it will produce worse policy.
The rational core is not Zuckerberg’s China argument. That argument is a lever, not a thesis. The rational core is the claim that premature regulation can freeze safety assessment into a form that is easy to pass but hard to trust. If the government writes a checklist, labs train toward the checklist. If the checklist is wrong, the industry has the illusion of oversight and the reality of gaming.
A pre-release certification scheme built on immature metrics does not reduce risk. It launders risk into compliance. The model that passes the test is not safer. It is better at passing the test. That is a different property and it is dangerous to confuse the two.
Sanders’s bill also carries a hidden anti-innovation bias. If the definition of prohibited capability is vague, the safest business decision is to do nothing ambitious. That is not precaution. That is freezing the frontier in favor of incumbents. A small lab with one extraordinary experiment will not take a twenty-year legal risk. The large lab with a compliance team can afford the test to prove its model is legal. The incumbents win again.
So the contrarian truth is that the worst regulation is not lenient regulation. The worst regulation is confident regulation built on unvalidated methods. That version gives society the appearance of control and the absence of control. Every audited protocol that collapsed after receiving a clean report was proof of that failure mode.
What an Actual Standard Would Require
If I were asked to build the missing audit infrastructure, I would start from a different place than Washington is starting. I would not begin with a regulator. I would begin with an invariant.
A model release should carry a claim that is falsifiable. The claim might be: under a defined threat model, the model cannot produce action X. It might be: the model cannot complete a specific harmful chain of operations unless a human with a named credential authorizes it. The claim has to bind the model’s outputs to observable system events. That means cryptographic logging, immutable audit trails, and deterministic verification of the governance layer around the model.
This is not a purely technical exercise. It is a legal architecture. The regulator should enforce the audit trail, not the vibes.
In securities regulation, the disclosure regime survives because the facts being disclosed can be checked. In AI, the system behavior is too complex to be captured by a single pre-release human-readable report. The report must be machine-readable, continuously updated, and independently replayable. That is the only kind of “disclosure” that matches the artifact.
The technology for this is not impossible. It is the same technology used for blockchain monitoring: signed events, chronological hashes, anomaly detection, and forensic reconstruction. The missing piece is not code. The missing piece is a juridical framework that treats AI behavior as a continuous process rather than a one-time movie review.
The Takeaway
The fight between Zuckerberg and Sanders is a fight about the kind of referee, not about whether the game needs one. Both sides accept the premise that someone will sit in the chair. Neither side is asking the question that matters before any referee is hired: what are the rules, and can they be checked against reality?
An agency with no testable standard is not a referee. It is a goalpost. A bill that bans an undefined threshold is not a safety policy. It is a press release with felony counts. A safety test that samples rare behavior and calls it clearance is not an audit. It is a signature.
I do not fix bugs. I reveal the truth you hid. The truth here is that the regulatory debate has skipped the engineering phase and moved directly to the enforcement phase. That inversion guarantees the next collapse will be explained as a failure of imagination. It will not be. It will be a failure of verification.
Hype burns hot; logic survives the cold burn. The cold logic of this moment is that neither speed nor safety can be legislated into existence without an instrument that both sides can inspect. Until that instrument is built, every new AI regulator is a warm body without a thermometer.
I have seen enough governance failures to recognize the smell. Every gas leak is a story of human greed. The 2026 AI regulatory showdown is not a story about machines. It is a story about humans fighting for control of a machine they do not yet know how to measure. The referee is coming. The question is whether the referee will be blind on one side or completely blind on both.