The Committee That Doesn't Exist Yet: On AI Governance and the Quiet Search for a Common Ledger

On a Tuesday that felt like every other Tuesday, a senator from New Mexico put his name behind a letter that asked the United States Senate to admit it doesn't know who is in charge.

Ben Ray Luján — Democrat, land of Los Alamos and Sandia, land of the nuclear scientists who taught America how to hold something dangerous without dropping it — urged Senate leadership to create a select committee on artificial intelligence. Not a task force. Not a working group that dissolves when the news cycle turns. A committee. A permanent room with a door and a nameplate and, presumably, jurisdiction.

The letter itself is four points long and contains no model weights, no compute thresholds, no benchmark tables. It is, by any technical reading, thin. And yet I have read it three times, because I recognize the shape of it. I have watched this shape appear before, in a different discipline, among people who were also trying to write the rules for a system that had already run ahead of them.

This is not a story about AI. It is a story about coordination — and coordination is the only problem that has ever mattered in decentralized systems. Every broken token taught me how to hold value. This letter is someone learning to count again.

Context: Fifty Rulebooks and No Referee

To understand what a select committee would actually be, you have to understand the mess it would be trying to clean up.

As of early 2025, the American AI industry operates under a patchwork of state-level rules that resemble nothing so much as an unsynchronized set of nodes. California's SB 1047 — vetoed, but only after pulling the entire national conversation into its gravity well. Colorado's AI Act, the first comprehensive state framework of its kind. A growing stack of disclosure requirements, algorithmic auditing mandates, and procurement conditions spread across a dozen legislatures, each drafting in isolation, each defining "high-risk" slightly differently.

For a startup building one model, this is not fifty markets. It is fifty compliance stacks. Fifty interpretations of what transparency means. Fifty answers to a question no engineer ever asked: who signs the form when the model misbehaves?

Meanwhile, the European Union has done the opposite. The AI Act exists. The AI Office exists to enforce it. Brussels wrote a single, vertically integrated regulatory organism and is phasing it in on a schedule. Love it or hate it, it is coherent. And in the language of distributed systems, coherence is a form of consensus — even a wrong consensus is more useful than a fragmented truth.

This is the context Luján is writing into. Not a philosophical debate about machine personhood. A plumbing problem. The US has delegated its AI rulebook to the lowest common denominator of a federation, and the federation keeps disagreeing with itself.

The request, then, is modest: give the Senate one table where the AI question lands, instead of scattering it across Commerce, Judiciary, Health-Education-Labor, Energy and Commerce, and Homeland Security — five committees, five appetites, five sets of donors and priorities. A select committee is an attempt to elect a leader.

Core: What Fragmentation Actually Costs

Let me be precise, because the word "fragmentation" gets thrown around like "liquidity" in a bull market.

Regulatory fragmentation is a tax, and it is paid in exactly the ways decentralized systems pay for uncappable state. When I audited Uniswap V2's contracts — three hundred hours, not for vulnerabilities but to understand the fairness embedded in immutability — what struck me was not the elegance of the code. It was the assumption beneath it: that a rule published once, visible to everyone, and enforced without discretion is cheaper than a rule interpreted fifty times. The whole point of a canonical contract is that it removes the question of which version of the truth you are running.

AI governance has no canonical contract. It has fifty forks, all live, none backward-compatible.

The real cost isn't paperwork. It's arbitrage. When standards diverge, capital and compute migrate toward the loosest jurisdiction — the same way liquidity mining APY, stripped of its incentives, reveals how few of those users were ever there for the protocol. Push a compliance burden past a certain threshold and you don't get better behavior. You get a shell headquarters in a friendly state and a research team somewhere else. You get, in other words, exactly the kind of regulatory tourism that a unified federal baseline is meant to foreclose.

And here is the part the letter does not say, because it doesn't have to: the two outcomes are not equally good. Federal unification can mean raising the floor or lowering it. Preemption — the doctrine that federal law displaces state law — is not automatically a gift to safety. It could freeze a stricter Colorado standard into irrelevance. The industry reads "consistency" and hears "cheaper." The safety community reads "consistency" and hears "weaker." Both readings are coherent. That ambiguity is the actual content of the debate, and the letter is silent on it.

What I keep returning to is the international dimension, because that is where the stakes get geometric. The EU has a referee. China operates a top-down framework that, whatever else you say about it, produces one signal instead of fifty. If the US cannot stand up a unified position, it forfeits the thing that matters more than models in the present era: the authority to write the standard others adopt. In a market defined by rules, the rule-writer is the real competitor. The model is downstream.

That is the insight buried under four policy points. The competition between the US, EU, and China is no longer primarily about who trains the best system. It is about who supplies the vocabulary — the definitions, the audit norms, the reporting formats — that everyone else must translate into. Regulatory coherence is a form of standardization power. Luján is, perhaps without saying so, asking for a pen.

Contrarian: This Is a Turf War Wearing a Safety Badge

Now the part that makes people uncomfortable.

Every governance structure is a coordination mechanism and a distribution of power. These are not separate facts. They are the same fact read from two angles.

A select committee on AI is, functionally, an institutional answer to a jurisdictional question: which Senate committees currently own AI, and which would have to give something up for a new one to exist? Right now the answer is five of them. Creating a single select committee is not just building a house. It is asking five other houses to redraw their property lines. That is why the letter is addressed to leadership — the only actors who can settle a boundary dispute between committees, because they control the assignments that determine whether a senator's years of work translate into anything.

So when I read the four points, I don't read them as a safety blueprint. I read them as a positioning document. Luján's state is home to national laboratories whose entire existence is a bet on federal science funding and federal compute. That is a legitimate reason to want a committee. It is also a reason that has nothing to do with model safety, and pretending otherwise is how the industry gets fooled. I have watched this exact move at a much smaller scale: communities that launch a "values-driven" governance token and, eighteen months later, discover the token distribution was the constitution all along. The charter is not the covenant. My code was the covenant, not just the contract.

The contrarian claim is this: the presence of a select committee tells you almost nothing about whether AI will be governed safely. Institutions can centralize authority without centralizing standards. A committee with subpoena power but no enforcement mandate produces hearings, and hearings are not audits. The AISI — the AI Safety Institute — already exists and sits with limited authority precisely because no one wanted to give it teeth. A new committee could be the thing that finally grants teeth, or the thing that absorbs the pressure for teeth into a room where testimony substitutes for action. The structure is neutral. The mandate is everything, and the letter doesn't specify one.

There is a quieter risk, and it lives in the crypto-native corner of this story. A unified federal AI posture will almost certainly be written with centralized actors in mind — large labs that can afford compliance departments. Decentralized AI projects, which route around exactly this kind of gatekeeping, will be the first casualties of a clean, legible standard. That is not necessarily wrong. Legibility has value. But it is worth naming, because the media outlets covering this letter serve an audience that built things precisely because the old coordination layer kept excluding them.

Takeaway: Build the Room Slow, Keep the Ledger Honest

I don't know whether the Senate will form this committee. I suspect it will not — not in this session, not with this leadership, not while AI remains a topic everyone wants to be seen legislating and no one wants to be the last one holding the pen.

But I know the direction of the current. Fragmented truth never beat a single ledger. The question is never whether to coordinate. It is whose version of the covenant gets written down first.

So here is what I will be watching — not the headline, but the mandate. Does the committee get subpoena power, enforcement authority, or merely a mailbox? Does it preempt Colorado and California, or ratify them? And does anyone, at any point, write a rule that treats a decentralized model as something other than an anomaly to be regulated away?

In the silence of the bear, we heard the truth: systems fail not when they are wrong, but when they cannot agree on what right means. The AI industry is loud right now. Which is exactly why I am listening for the committee that doesn't exist yet — and for whoever, in the end, is allowed to sign its name.