The date on the FTC docket says 2026. The FRONTIER Act carries a July 23, 2026 introduction date. The Capitol Hill gathering at the center of this story happened September 15–16, 2025. The Truth Social post arrived September 19. The timeline does not reconcile. That is the first finding worth reporting.
A policy narrative that cannot keep its own dates straight deserves a second pass. Data leaves footprints; hype leaves only dust. This footprint is smudged.
The gathering assembled a strange political coalition: three legislators and a populist agitator. Bernie Sanders, pushing a blanket ban on "artificial superintelligence." Ted Lieu and Jerry Moran, who want emergency kill-switch authority in the hands of the DHS Secretary. Josh Hawley and Dick Durbin, who want AI classified as a product subject to federal product liability. And Steve Bannon, whose diagnosis is blunter: tech companies have spent a decade externalizing risk onto the public.
The coalition calls itself Pro-Human. Its shared premise: AI accountability should not be optional. Its shared problem: nothing else is shared.
Four legislative paths emerged from the gathering. On paper, all respond to executive overreach — Trump's newly announced "AI Force" and an unnamed, undefined AI Czar. In practice, they encode four mutually exclusive philosophies of risk allocation. That mutual exclusivity is the story.
The Threshold Arbitrage Problem
The FRONTIER Act anchors its regulatory architecture to a single technical parameter: 10^26 FLOPs of training compute. Three obligation tiers follow. Small-scale training warrants model cards. Large-scale training activates a catastrophic risk framework. Ultra-large-scale training demands continuous audits by licensed independent verification bodies.
The number is measurable. It is also a weak proxy for capability.
Distillation, synthetic data, and test-time compute all produce high-capability systems at a fraction of the training footprint. The threshold creates a built-in arbitrage incentive: obtain the capability without paying the compute, and you escape the regulatory tier. This is threshold arbitrage, and the bill's design makes it systematic. Code is law only until someone finds the loophole. Here the loophole is algorithmic efficiency itself.
My forensic work confirms this pattern with grim reliability. Projects optimize for the question they are asked, not for the risk they actually carry. In 2022, I found an integer overflow in the withdrawal function of a $12 million cross-chain bridge that skipped third-party review. The team had a mainnet deadline. They did not have a working understanding of their own code. Regulators who build thresholds are building the same false comfort. Audits check syntax; journalists check motive. The motive behind a 10^26 FLOPs line deserves more scrutiny than the number itself receives.
The Definition Vacuum
The Sanders path proposes permanently prohibiting the development or deployment of artificial superintelligence. No capability threshold. No evaluation suite. No certification mechanism. A ban without an operational definition of its target is a press release wearing legislative clothing.
The bill had not even been formally introduced at the time of the gathering. That makes it a media event, not a legislative event. Media events do not create liability.
The Kill Switch That Cannot Kill
The Lieu–Moran path grants the DHS Secretary emergency authority to order AI systems slowed or shut down. Technically plausible for closed, API-hosted models. Completely inapplicable to open-weight models. Llama, DeepSeek, Mistral — once weights circulate, no off switch exists anywhere on the planet.
The kill-switch requirement operates, in practice, as a unilateral constraint on closed-source vendors and a structural subsidy to the open ecosystem. Regulators never intend arbitrage. They manufacture it anyway.
Where the Definitions Come From
The 10^26 FLOPs threshold is not original to this bill. It descends from Executive Order 14110's reporting requirement — a disclosure line quietly upgraded into a regulatory boundary. That is a stealth escalation of regulatory intensity, and the industry has not priced it.
Beneath every whitepaper lies a buried intent. For this coalition, the question is whether the intent buried inside the thresholds belongs to the public or to the labs that supplied the technical language. Vendors have a demonstrated incentive to position the line above their own current compute scale. Regulatory capture via definition requires no bribes and no lobbying — only the privilege of being the source of the numbers.
All four paths regulate the training side of AI. None address inference scale, deployment breadth, or agent autonomy. The architecture is out of phase with capability diffusion, structurally.
The Business Friction Nobody is Pricing
Enterprises deploying agents that make real purchases, real bookings, and real communication decisions face a legal vacuum. Four coexisting paths mean no coherent federal framework. Liability cannot be priced. Insurance cannot be underwritten. Under the Hawley–Durbin path, AI is a product, Section 230 immunity is stripped, and developers and deployers stand as co-defendants. That exposure cannot be contracted away.
State-level fragmentation is already arriving. Connecticut's AI liability regime takes effect October 1. Maryland's algorithmic pricing statute takes effect the same day. The FTC is collecting comments on AI-driven personalized pricing. Compliance costs are scaling linearly per state, and no federal bill has passed to stop it.
The counterintuitive detail the headlines miss: the Trahan path's preemption clause is the most industry-friendly provision in any of the four frameworks. It blocks the 50-state patchwork before it congeals. A regulatory bill containing preemption is, for enterprise players, a compliance cost reduction disguised as oversight.
The Bull Case the Consensus Misses
The reflexive read is that stricter AI liability hurts the incumbents. The evidence cuts the other way.
If the Trahan framework survives, audit and verification costs become fixed compliance line items. Large players absorb them. Small players and open-source projects hit barriers to entry. Market concentration rises. The liability firewall, in its operationalizable form, is a moat-digging instrument.
Open-source models collect a regulatory arbitrage premium because enforcement is infeasible at scale. Closed vendors carry the audit burden; open weights do not. In my 2026 review of three "autonomous economic agent" protocols, every one turned out to be a script wrapped around a centralized API. The founders called it decentralized intelligence. It was a cron job with venture funding. That lesson transfers directly: when regulation fastens onto centralized control points, it rewards the unregisterable. Markets price this asymmetry faster than regulators perceive it.
The compliance industry is the quiet structural winner. The FRONTIER Act's licensed verification bodies resemble the accounting majors that minted fortunes post-SOX. AI audit, certification, and compute-verification services are a predictable revenue stream wearing a public-interest uniform.
What Actually Comes Next
The coalition is politically heterodox: progressive, moderate, libertarian, populist. That breadth is a weakness wearing strength's clothing. Coalitions built on a single point of consensus disintegrate at the clause level. The four paths are legally incompatible. Preemption and state liability authority cannot coexist. The Sanders ban is absolutist. The kill-switch is untethered from operational reality.
Meanwhile, the executive branch does not need sixty votes to act. The AI Force is being assembled. The AI Czar, whoever it turns out to be, will hold levers requiring no legislation. Institutional design favors the executive. The default outcome is not a firewall. The default outcome is a regulatory vacuum.
Deployers of autonomous agents should build for that vacuum. Assume no federal liability protection in any realistic planning horizon. Design internal accountability structures. Price in state-level fragmentation.
Watch for three signals: whether Sanders' bill receives a legislative number within the next month, whether the AI Czar is actually nominated, and whether the FRONTIER Act secures a committee markup. None of the rhetoric matters until one of those happens.
The window is not closing. It is calcifying into a vacuum.
Truth is not distributed; it is discovered. Someone needs to publish the audited findings.