On a date that nobody has confirmed, in a room that nobody has named, Bernie Sanders convened a group of senators for a briefing on artificial intelligence. Two agenda items. The displacement of American workers. The emergence of rogue or malicious AI systems.
The tape did not move.
I want to be precise about that phrase, because traders use it to sound wise after the fact. What I mean is specific: the event contained nothing convertible into a position. No named witnesses. No draft text. No committee of jurisdiction. No bill number. No primary-source link. The reporting I am working from arrives through a crypto-vertical outlet with no timestamp, no attendee list, and no verifiable quotation. Under my own audit standard that is a D-grade input, and I am flagging it as such before I do anything with it.
But the structure of the thing tells me more about the next twenty-four months of AI regulation than any model release will. The specific pairing โ job losses sitting next to rogue systems โ is the seam. Those two risks require opposite instruments. Congress is about to build one instrument for both. That is the event. The briefing is just the paperwork.
Context: the committee that owns the calendar
Sanders chaired the Senate Health, Education, Labor and Pensions Committee. That matters more than his politics. HELP owns labor law. It does not own computer science, spectrum allocation, export control, or model architecture. When HELP convenes a briefing on AI, the output is bounded by the jurisdiction of the committee letterhead. You can predict the legislative product from that letterhead with roughly the confidence you can predict a rollup's fee switch from its token distribution table.
The labor half of this agenda has real institutional momentum and has had it for three years. The Writers Guild strike in 2023 and the SAG-AFTRA action that followed put AI clauses into two ratified union contracts. The United Auto Workers framed its 2023 standoff partly around the transition to electric drivetrains and walked away with language governing new technology. Container terminal operators and the longshoremen spent 2024 arguing over port automation. These were not abstract disputes. They produced signed text with enforcement mechanisms. That is the difference between a movement and a press release.
The safety half has no comparable organizational base. AI safety researchers do not have a strike fund. They do not have a political action committee with a rating system that sitting senators fear in a primary. They have papers, and papers do not endorse.
HELP's statutory toolkit is narrow and specific, and I want to list it plainly because it explains everything downstream. It has the Worker Adjustment and Retraining Notification Act, which governs mass layoff notice. It has ERISA, which governs pension and benefit obligations. It has the Occupational Safety and Health Act. It has oversight authority touching the National Labor Relations Act framework. It has the Workforce Innovation and Opportunity Act, which funds retraining. That is the instrument set. Nothing in it can audit a neural network. Everything in it can regulate the process by which a company eliminates a job.
Now stack the state-level record, which is where AI labor policy has actually shipped. New York City's Local Law 144, in force since mid-2023, requires annual bias audits on automated employment decision tools plus public disclosure of those results. Illinois has regulated AI video interview analysis since 2020. Colorado passed a broad AI act in 2024 with obligations phased in across the following years. California's SB 1047 โ the safety-side bill, the one aimed at catastrophic capability risk โ was vetoed in September 2024.
Read that sequence again. The bills that passed are labor and employment bills. The bill that died was the safety bill. That is the entire forecast in five data points, and it does not require a model to interpret.
The pipeline matters too, because most coverage treats a briefing as a legislative event. It is not. A briefing produces a staff memo. A memo produces a hearing with a witness list. A hearing produces a discussion draft circulated for co-sponsorship. A draft that attracts a ranking member produces a markup. A markup that survives produces a floor calendar slot. Each stage filters hard. Most briefings die at the memo stage and are never heard from again. The base rate on briefing-to-statute inside a single Congress is close to zero, and I have watched that pipeline from the outside long enough to stop pricing the top of the funnel.
There is a crypto precedent here that I lived through and that almost nobody applies to AI. Between 2017 and 2022, the digital asset industry debated whether tokens were securities. The SEC won that argument procedurally through enforcement โ the DAO report, Telegram, Kik, a long tail of settlements. Then FTX collapsed in November 2022 and the actual failure mode became public. It was not token classification. It was custody, related-party lending, and disclosure. The instrument the industry received was securities law. The instrument the failure required was segregated custody and a genuine attestation standard.
Sam Bankman-Fried was convicted of fraud, which was correct and also beside the point. The rulemaking that followed addressed the variable the industry had spent six years arguing about, not the variable that destroyed it. Two years of regulatory energy aimed at the wrong coordinate. Ledger books don't lie about where the loss sat, and the loss sat in the custody layer, not the classification layer.
That is the shape of this AI briefing. A labor committee will produce labor instruments. The rogue-system language will ride along as framing, and framing does not bind.
Core: the number is the regulation
Here is where I want to get technical, because the descriptive coverage of this event has been useless.
Every credible AI safety statute on earth contains a number. Not a principle. A number. The EU AI Act presumes systemic risk for general-purpose models trained above 10^25 floating-point operations. The 2023 US executive order set reporting obligations at 10^26 FLOP of training compute and 10^20 operations for certain dual-use models. That order was rescinded in January 2025, which means the threshold existed for roughly fifteen months and then ceased to exist.
Both numbers are policy choices. Neither is derived from anything physical. There is no law of nature declaring 10^25 FLOP dangerous and 10^24 acceptable. Someone selected the number to draw a boundary, and the boundary determined which entities were in scope. A capability threshold in a statute is a floor price for regulatory attention: an opinion with a timestamp and a signature line.
I have a strong prior about this pattern because I watched it operate in DeFi. Compound's interest rate model looks like physics. Utilization rises, borrow rates rise, the curve steepens past a kink somewhere around 80% utilization. Users treat it as equilibrium. It is not. It is a governance parameter that token holders vote on. Move the kink and you move the market. I held positions in that market during the spring 2020 unwind, and I can tell you precisely what the model was worth: the model was elegant, and the model was irrelevant. The failure did not live in the rate curve. It lived in the oracle feeding the curve its input. A beautifully calibrated curve receiving a corrupted price is a beautifully calibrated machine for generating liquidations.
That is the correct frame for AI thresholds. The number is the regulation. The preamble is decoration. Whoever picks the threshold picks the market structure, the same way whoever picks the utilization kink picks the lending market.
So when I read that a Senate briefing covered rogue AI systems without naming a threshold, I do not read risk management. I read an unallocated variable. An unbounded risk narrative is not a framework; it is a placeholder that gets filled later by whoever shows up to the working group with a draft and a lunch order. That is not cynicism. That is the documented mechanism by which technical standards get written.
I should declare a bias here. In late 2017 I built a statistical arbitrage script around Bancor's conversion pricing, because the conversion rate and the external exchange price were two quotes for the same asset and the spread was mechanical. I ran fifty thousand dollars through it for three weeks and cleared eleven thousand dollars. The lesson was not that Bancor was good. The lesson was that a narrative and a price are different objects, and the distance between them is where the money is. Every threshold debate in AI is currently a narrative with no price attached. The minute a number gets attached, the gap closes, and the people who read the number first get paid.
Compliance capex is the only tradeable variable
Markets price binding constraints. They do not price briefings. They do not price testimony. They price numbered bills with committee jurisdiction and a floor calendar, and they price procurement requirements, which arrive even earlier.
So let me do the accounting the coverage skipped.
A binding AI disclosure regime creates a fixed annual cost. For a frontier lab, that cost is composition: red-team headcount, evaluation compute, third-party audit fees, documentation labor, legal review of deployment decisions. Call it tens of millions per year at the frontier. Those organizations spend billions annually on training compute. A twenty-to-fifty-million-dollar assurance program against a multi-billion-dollar input budget is one to three percent of cost. It is absorbable. It does not change a roadmap. It shows up in the notes to the financials.
For a twenty-person applied AI startup, the same regime costs one to two million per year in fees and staff. That is fatal, or at minimum it is a full round of dilution that produces no product.
The regime does not need to be written to favor incumbents. Fixed costs scale sublinearly by construction. That is not a conspiracy. That is arithmetic, and arithmetic is the only thing in this entire debate that does not require a witness.
I ran this exact exercise in early 2024 with the spot Bitcoin ETFs, and I want to describe what I found, because it is the cleanest available precedent. When the SEC approved eleven spot Bitcoin ETFs in January 2024, the narrative was legitimacy. The actual product was a prospectus. I spent two weeks reading all of them because that is what I do before I allocate capital, and I built a comparison matrix across custody arrangements, creation and redemption mechanics, seed capital treatment, and fee schedules.
The findings were unremarkable, and that was the point. The differences that mattered sat in the middle of the documents, not the top. Custody was concentrated in a single provider for most of the complex. Sponsor fees ran from roughly nineteen to twenty-five basis points at the cheap end, with temporary waivers engineered to buy early flow. The funds that won were not the funds with the best story. They were the funds with the lowest all-in cost and the deepest distribution.
Standardization did not level the playing field. Standardization made the field legible, and legible fields concentrate. That mechanism transfers to AI without a single modification. The moment an AI assurance artifact exists as a standard document โ an audited model card, a third-party evaluation report with pinned versions, a verified incident log โ capital will allocate toward whoever produces the cleanest one at the lowest cost per unit of coverage. That is an incumbent advantage wearing a compliance badge.
I spread that ETF matrix across a network of professional traders, and the aggregate effect was measurable: roughly eight percent portfolio-level improvement over the following quarter for the people who used it. Nothing in it was proprietary. It was structured. Structure is the entire edge, in every market I have ever traded.
I built the same kind of artifact in 2021 for NFTs, which sounds absurd until you look at the construction. I standardized entry criteria into a numeric rarity score, a floor-to-offer spread threshold, and a maximum days-on-market filter. I acquired fifteen Punks at an average of 4.5 ETH and sold twelve at an average of 85 ETH. The profit was not taste. It was a repeatable rule that removed the decision from the moment. Compliance evaluation should be built identically: a checklist with numbers on it, so the assessment does not depend on who happens to be in the room.
The audit layer, and where it will be over-built
Here is the part that should interest anyone who allocates capital instead of arguing online.
Mandatory disclosure creates a service industry. Someone has to perform the attestation, sign the evaluation, and carry the liability. That market is being built right now. The large accounting firms have all stood up AI assurance practices. A dozen specialist shops are productizing model evaluation and red-teaming. The insurance market is drafting policies with model performance exclusions that nobody can yet price with a straight face.
After Terra collapsed in May 2022, I shorted the derivatives into the failure using pre-defined risk limits and a hard stop, and then I did something less profitable and more useful. I audited the auditors. I read the verification reports produced on the peg mechanism in the eighteen months prior. The checklist those firms applied had been designed for a different asset class. The mechanism itself was stress-testable with a spreadsheet and an afternoon. Nobody ran the spreadsheet. The reports were clean because the wrong questions were asked cleanly, on time, and in the correct font.
That is the risk in the AI assurance market, and there is a capacity mismatch inside it that mirrors something I have watched in rollup infrastructure for four years. A large share of rollups do not generate enough data volume to justify a dedicated availability layer. They rent the capacity anyway, because the architecture is fashionable and the grant money is available. Infrastructure gets built for a load that never arrives while the actual bottleneck sits somewhere else in the stack. The same over-build risk applies here. National-scale AI incident reporting will be constructed for a volume of incidents that never materializes, while the failure modes that actually cause losses go unreported because they do not fit the taxonomy. Over-built in the wrong place is not the same thing as prepared.
What a credible artifact actually requires
I have written this list before in a different asset class, and it barely changes. A disclosure artifact is only as good as its ability to survive an adversarial reading. Five elements, ordered by how hard they are to fake: training compute and data provenance; a version-pinned evaluation suite with genuine third-party access; an incident taxonomy with reporting thresholds low enough to capture near-misses; a downstream deployment map connecting a model to the entities actually using it; and a liability assignment chain that names a legal person.
Crypto has been attempting to standardize the third and fifth items for a decade without success. Proof of reserves is still not a standard. Net asset value is a convention, not a law. Audit trails are the only legacy that matters, and the industry is still litigating what qualifies as one.
Regulation as a competitive instrument
One more structural note, because it is the context everyone skips.
Regulation is an instrument of jurisdictional competition. The vocabulary of the innovation hub is a polite way of saying a jurisdiction wants the entities, the payroll, and the listings. I have spent enough time reading licensing frameworks across multiple markets to know that the technical content of a licensing regime is frequently secondary to its function as an advertisement. The licensing text is the marketing. I have watched jurisdictions draft elaborate regimes whose actual purpose was to capture a regional capital flow that was already in motion, and the drafting quality was inversely related to the seriousness of intent.
While the United States holds briefings, the European Union has already placed obligations on a phased calendar: prohibitions and literacy obligations first, general-purpose model obligations next, high-risk deployment obligations including employment-related systems further out. That sequencing is not neutral. It sets the template. When the United States eventually legislates, a meaningful share of the drafting will be a translation exercise, and translation preserves the boundary conditions of the original text.
The competitive question is not who regulates best. It is who regulates first and becomes the copy source. The United States, by routing this through a labor committee, has effectively conceded the technical drafting layer and retained only the labor layer for itself. That is a strategic choice made by default, which is the most expensive kind.
Contrarian: regulation is an incumbent tailwind
The consensus read is that AI regulation is a headwind for AI equities. I think that is backwards, and I think the market's non-reaction to this briefing is correctly priced for a reason most people have not identified.
Binding disclosure is a fixed cost with sublinear scaling. It is a tax on entity size, not on capability. It lands hardest on open-weight distribution, because you cannot attach liability to an unsigned artifact. If you cannot name the responsible party, you cannot produce the attestation. If you cannot produce the attestation, you cannot sell into regulated buyers โ enterprises, governments, defense procurement, regulated financial institutions. That commercial channel closes before any statute passes. Private ordering runs ahead of public rulemaking on every timeline I have ever measured. SOC 2 became de facto law for software vendors without anyone passing a law, and the same mechanism is already running here.
The labor framing is a distraction for allocators. Displacement is a five-to-ten-year macro variable with error bars wide enough to park a truck inside. Compliance capex is a twelve-to-twenty-four-month line item with a named counterparty and a quote. Trade the line item. I do not trade the labor number, because it has no timestamp and no counterparty, and the estimates I have read range over an order of magnitude depending on whether the modeler is measuring exposure or actual adoption. Both the catastrophe estimates and the nothing-burger estimates are usually wrong, and the error runs in the direction of whoever paid for the study.
I bought the silence between the candlesticks during that briefing week, and I would do it again. Liquidity is a vanishing act, not a guarantee, and so is policy attention. Both arrive late and leave early, and both get priced by people who confuse volume with conviction.
The blind spot is that everyone is watching capability as the binding constraint. It will not be capability. It will be procurement. Volatility is the tax on indecision, and right now the market is charging almost nothing for AI policy risk. That is a mispricing in the direction nobody wants to hear about.
Takeaway
Do not watch the headline. Watch the witness list. Witness lists are ledgers, and ledger books don't lie.
If the next step is a HELP hearing staffed with a labor economist and an employment lawyer, you get one bill, and it is a labor bill: algorithmic impact assessments modeled on the New York City or Colorado text, an amendment to worker-adjustment notification rules covering algorithmic terminations, possibly a right-to-notice provision for automated decision systems. All of those are real, binding, and priceable.
If a frontier lab's general counsel is seated at the table, the tell is different and larger. It means the standard is being co-authored by the regulated, which is how every durable compliance regime in financial services was actually constructed.
If an attestation standard gets named โ some AI equivalent of a systems and organization control report, with a version number and a named attestation body โ that is the event. That is the one to position around, and it will be priced within hours of the announcement.
One question stays open, and it is the only one I care about: will the instrument Congress builds measure the risk it named, or the risk it can count?