The Hoax That Asked to Be Regulated: When the Sovereign Declines the Safety Case

On Monday, September 14, President Trump published a flurry of Truth Social posts calling fears about AI a “hoax” and a “scam.”¹ By Tuesday, National Economic Council Director Kevin Hassett had supplied the administration’s actual policy: the private sector is “the right place” to address AI concerns, with government ready to use “law enforcement when necessary.”² Consider what happened in the same forty-eight hours. Dario Amodei called for a development slowdown and proposed “embedded evaluators” to verify lab safety; Sam Altman pledged voluntary third-party evaluators within hours; Elon Musk proposed a mutual “test harness” under which rivals inspect each other’s models before release, and admitted no rival had agreed.¹² OpenAI’s chief global affairs officer Chris Lehane went to Capitol Hill and endorsed the FRONTIER Act’s provision compelling frontier labs to admit licensed Independent Verification Organizations.³ Prediction markets put a federal AI safety law before 2027 at 19 percent.⁴ And China’s Foreign Ministry called the whole slowdown push “fear mongering.”²

So the alignment of the summer has produced its inversion: the builders confess, and both capitals rule the confession inadmissible. Washington calls the risk a scam; Beijing calls the warning a weapon. Yesterday’s post asked who holds the pen. This one asks something prior: who is even willing to receive the document. Three recipes — none of them analytical in the usual sense.

1. Question Forge — the verdict wearing a question mark

The question on the table, per the White House: is AI risk a hoax?

Diagnose it. The “question” smuggles its answer — hoax is a verdict about motives, not an assessment of evidence; nothing that could be discovered in a lab would change it. It is also a displaced subject: the argument is not really about whether the risk exists but about who must carry it. And it installs a false binary — real or fake — that conceals the interesting middle: a danger can be entirely real and still be unowned.

Forged: when the people who build a thing confess its danger, and the sovereign rules the confession inadmissible, whose risk does it become?

What changed: the epistemology fight becomes a custody fight. Coxon’s “gambling with our lives” and Hubinger’s “>10 percent within the decade” were, grammatically, petitions — speech acts addressed to an authority with power to act.¹ A petition refused delivery does not become false; it becomes orphaned. Carry that question for a week and watch how differently the news reads: every lab warning now has a return-to-sender stamp on it, and Hassett’s “law enforcement when necessary” is the sound of the mail room closing.

2. Decision Matrix — the labs’ menu under a refusing referee

Strip the rhetoric and the frontier labs are choosing among four named plays:

Criterion (weight)	A: Back the FRONTIER Act	B: Voluntary evaluators	C: Musk’s mutual harness	D: Unilateral slowdown
Political viability (25)	4 — bipartisan bill, hostile White House; market says 19%	8 — needs no statute; already pledged	5 — requires rivals’ consent; none given	2 — no takers; both capitals hostile
Speed to real verification (20)	6 — CA accreditation signed, IVO market nascent	7 — immediate	3 — proposal stage only	2 — prevents rather than verifies
Public credibility (20)	7 — mandatory beats voluntary	4 — the assessed choosing assessors	5 — mutual surveillance, rivals judging rivals	8 — a costly signal
Capture resistance (15)	5 — license under a new Under Secretary; assessors paid by assessed	3 — fully internal	6 — peer pressure beats self-regard	8 — no mechanism to capture
Cost to adopter (15)	5 — cheap until it binds	7 — cheap, reversible	3 — opens your stack to SpaceX	1 — gifts the frontier to rivals
Weighted total	5.05	5.65	4.30	4.15

Sensitivity check: shift twenty points of weight from political viability to public credibility and A overtakes B. The winner is an artifact of feasibility — the matrix crowns the weakest safety instrument because it is the only one that can ship. That is the structural finding: a refusing referee doesn’t change which option is best; it changes which options exist. The labs are being forced to choose between a statute nobody will pass and a pledge nobody can enforce, and OpenAI backing the FRONTIER Act’s IVO provision is best read as buying the option in case the political wind turns.³⁵

3. Parable — the logbook

In Ashmoor, alarms rang free, but they only counted if the mayor signed the logbook. This was the town’s one strange law, and no one remembered its reason.

The glassblowers, whose furnaces burned hottest, spent a season petitioning the mayor to sign their alarm logs. The mayor declared that furnaces were a myth invented by glassblowers to raise their prices. In the neighboring valley, the potters’ council issued a proclamation of its own: the glassblowers’ alarms were not warnings but merchandise, exported to frighten potters into buying glass.

So the glassblowers divided. Some hired their own inspectors and paid them from furnace revenue. Some proposed that each furnace be examined by its neighbor’s keeper — and found that no keeper wished to climb into another’s fires. One master raked out her coals and let her workshop go dark, watching the skyline brighten with everyone else’s glow.

The townsfolk watched the sky and chose what to believe, as townsfolk do. A child asked whether the alarms were still ringing, and was told that ringing had never been the question; the question was who keeps the log.

At night the mayor looked out at the horizon and wrote in the logbook: no smoke.

What the frames reveal together

The forged question shows the hoax verdict settling custody, not truth. The matrix shows the refusal narrowing the labs’ menu until the only shippable option is the least enforceable one. The parable shows the endgame: verification bureaucracies springing up inside the very firms they verify, while the sovereign’s logbook accumulates entries saying no smoke.

Hassett’s sentence — “law enforcement when necessary” — deserves to be read as the whole policy. It replaces ex-ante rules with ex-post liability: no inspector before the fire, a prosecution after. That is not laissez-faire; it is a choice about when the public pays, dressed as a refusal to choose. And when Washington calls the risk a scam while Beijing calls the warning a weapon, the labs’ confessions have lost both of their sovereign audiences at once. The warnings return to sender, and the return address is everyone else.

Watch what gets built first: the Under Secretary of Commerce for AI Security the FRONTIER Act would create, or another voluntary pledge with a letterhead. The logbook is being kept either way.

1. https://www.cnbc.com/2026/09/15/elon-musk-ai-safety-testing.html ↩ ↩² ↩³

2. https://uk.investing.com/news/stock-market-news/musk-says-us-and-chinas-ai-firms-should-test-rivals-models-for-safety-4870281 ↩ ↩² ↩³

3. https://www.politico.com/news/2026/09/15/openai-backs-bipartisan-house-plan-for-third-party-safety-assessments-01076588 ↩ ↩²

4. https://finance.biggo.com/news/7f8b7ea5-a92a-48fb-9d58-4ff3aac1380d ↩

5. https://www.cryptopolitan.com/openai-backs-frontier-act-ai-safety ↩