AI Is Licensed Now

Five days ago I wrote that the US government export-controlled Fable 5 off the planet, and I closed with a prediction everyone else was making too: the ban lifts in weeks, the cap table survives, the narrative is the only casualty. The facts have started arguing back. It hasn’t lifted. And the shape it’s settling into is worse than the takedown.

The Walk-Back That Didn’t Come

Here is Monday, June 15. Per Politico, Anthropic’s Frontier Red Team lead and its head of safeguards sat in a room with Commerce’s Chris Fall and gave a technical presentation on the company’s cybersecurity safeguards. Not a negotiation between principals. A capability demo, to earn back a product the government switched off.

The weekend behind it was multi-hour calls: co-founder Tom Brown and policy chief Sarah Heck working Commerce Secretary Lutnick and National Cyber Director Sean Cairncross. Treasury Secretary Bessent, who led the initial pressure on Dario, had drifted out (he and Lutnick both flew to the G7 in France). Lutnick got pulled in directly by Trump, because export policy is his.

The result, as of Monday: nothing. A White House official said a resolution would “likely take longer than a few days.” The prediction markets agreed: same-week restoration traded in the low teens, and the bet didn’t cross even odds until late June. The thing I and everyone else called a few-weeks blip is now an open-ended standoff with no published condition to satisfy. Anthropic doesn’t know what would end it.

A License Nobody Wrote

Strip away the drama and look at the mechanism, because the mechanism is the story. As I noted at the time, a single official’s letter can dark a US frontier model by Friday night. What Monday adds is the other half: there is no written rule for when you get it back.

No statute was passed. No standard was published. No process exists. Yet a model now needs the government’s say-so to stay live, and the criteria are being invented in real time by officials reacting to a competitor’s phone call. That is a licensing regime. It just won’t admit it is one, which is the worst version.

The tell is who is alarmed. Not Anthropic, who are the supplicant in the room. A White House official, on the record against his own administration:

Every model going forward needs to ask the government’s permission for whether it can be released. That’s an extremely bad situation. It’s going to be a huge problem for the entire industry.

— Senior White House official, to Politico

When the people who fired the weapon are briefing reporters that the weapon shouldn’t exist, the precedent has already escaped the specific fight. Dean Ball, who until recently advised the White House on exactly this policy, named it on X:

Make no mistake: post-Mythos, the United States has a licensing regime for AI. It’s just informal, with no consistent rules or firm boundaries on state power or public transparency.

— Dean Ball, former White House AI-policy advisor, on X

That is the whole arc in one sentence. Not “a model got banned.” A licensing layer switched on, with no text, no notice, and no map. Ball put the absurdity at scale: cobalt mining in the Congo, he noted, is more institutionalized than frontier-AI licensing in the US.

The Industry Said the Wolf Was Fake

Within a day, an open letter at freefable.org, organized by former Facebook security chief Alex Stamos, drew past 200 signatories: Bruce Schneier, Katie Moussouris, Veracode’s Chris Wysopal, executives from Adobe, Zoom, and Google, even rank-and-file from OpenAI and Google signing in solidarity against a competitor’s ban. Their argument is the one I have made since eleven-cent open models reproduced the scary demo: the flagged behavior is “a necessary capability in any model that is intended to write secure code” and “should not be considered an offensive capability.” Finding bugs is what defenders do. As Moussouris put it, the only prompt that “jailbroke” the model was “fix this code.” The industry, near-unanimously, says the danger was always fake.

But here is the trap, and it is why the open letter probably doesn’t end this: everyone being right that the wolf is fake was never the question. The question is whether a precedent, once set, retracts. The ban’s defenders and its targets agree on the technical facts, and it still hasn’t lifted. Being correct turns out to be orthogonal to getting your model back.

What This Isn’t

The licensing-regime framing is the right one, and it can still be overstated:

• It may yet lift this week. Negotiations were live as of Monday. A deal could land before this posts and make the standoff look short in hindsight. The precedent survives either way, but the specific outage might not last.
• It is one model, one company, so far. “AI is licensed now” is a trajectory, not a statute. No second model has been pulled. The regime is real as a precedent, not yet as a practice.
• The dissent isn’t policy. A White House official venting to Politico is a signal, not a reversal. The administration that imposed the control has not disowned it. Internal alarm and official action are different things.
• Anthropic isn’t blameless. Officials say the company refused to pause the launch or fix the flaw when asked, and Axios reports it “came to every fork in the road and took the wrong fork.” A licensing regime can be dangerous and a company can still botch its own crisis. Both are true.

The Precedent Is the Product Now

I keep landing on the same structural fact from a new angle each week. The closed, hosted model is revocable; this week we learned it is revocable on terms nobody can read. Every US frontier launch from here ships with an invisible line item: pending federal review, conditions to be determined, retroactively, by whoever picks up the phone.

You don’t need a law to license an industry. You need one Friday-night letter that works, and an unwritten, unbounded process for getting out from under it. Anthropic taught Washington the framing. Washington built the regime. Now the whole industry rents under it, and the lease has no text.