No company should have its lawful work switched off by a phone call.
Friends,
On June 12th, the federal government gave Anthropic ninety minutes to shut down two of the most capable AI models in the world. No law on the books. No notice. No evidence. No hearing. Just a letter from the Bureau of Industry and Security at the U.S. Department of Commerce – and by 10 p.m. the models were dark.1
It’s worth understanding how that decision got made. Over a single evening, a handful of Anthropic’s own competitors got senior officials on the phone, warning that the model could be talked into reviewing software for security flaws2 – something nearly every AI model on the market will do, and something the model was built to do for defenders.3 No study, no outside experts, no chance for anyone to respond. By the next night the letter had gone out. That was the whole of it: a few calls from interested parties, and a letter to make it official.
One of those models was at the center of what we’re building at Legion. Legion is a legal-tech company: we use AI to draft fully formatted pleadings, discovery, and motions for litigators. The premise is simple – one good lawyer should be able to take on the many, and our tools make that possible. By enabling attorneys to draft in minutes what used to take days, we help attorneys deliver better outcomes for clients at a price ordinary people can afford. That’s what the June 12th shutdown reached into.
Given everything we have riding on this, you might expect me to argue that AI carries no risk and the government should keep its hands off entirely. I’m not.
The risk is hard to measure – but that is exactly why you need process. You bring in the experts. You sit down. You transparently build a way to assess the danger before you act on it.
What happened here was the opposite of that. No experts. No measurement. No transparency. Just a decree from on high, aimed at one company’s two models – while the same capability the government pointed to stays available in competing products it didn’t touch.4
Laws and procedures are what stand between us and power exercised on a whim, against whoever’s inconvenient. And the record here is hard to read any other way. Right before the shutdown, the President signed an executive order that expressly provided that nothing in it authorizes a mandatory licensing, preclearance, or permitting requirement for releasing AI models.5 Ten days later, the government did exactly that – to one company. The same company a court had already found this administration retaliated against once before.6
When the people in charge claim a power they themselves disclaimed in writing, and aim it at a single target, this is discretion without limit and without accountability – and that’s a threat to our constitutional freedoms.
It’s also a threat to America’s ability to lead in AI. Businesses don’t fear hard rules; they fear chaos. Clear rules and predictable process are what make it safe to build here. Rules made on a whim – or worse, ones that hamstring the people doing the building – do nothing but set the country back in a race other nations are sprinting to win.
What I believe is that this is government overreach. No company, and no person, should have their lawful work switched off by a phone call.
Legion exists to power the David-versus-Goliath fights so the few could take on the many.
This time, we’re the few.
And we’re suing.
Sincerely,
Arthur E. Rothrock, Esq.
Founder & CEO, Legion
Sources & notes
- 1. The June 12 directive, the ninety-minute deadline, the Commerce/BIS letter, and the worldwide shutdown: Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5”; Al Jazeera, “US orders Anthropic to disable AI models for all foreign nationals”. ↩
- 2. That the shutdown was set in motion by calls from Anthropic’s competitors to senior officials: TechCrunch, “Amazon CEO reportedly raised Anthropic model concerns before government crackdown”. ↩
- 3. That the flagged code-review capability is common to modern models and used by defenders: Anthropic statement (the capability “is widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe”); CSIS analysis (the vulnerability “is inherent to all modern language models”). ↩
- 4. That the same capability remains available in competing products the directive left untouched: Anthropic statement (naming OpenAI’s GPT-5.5); Al Jazeera, “US asks Anthropic to block global access to top AI models: Why it matters”. ↩
- 5. The June 2, 2026 executive order disclaiming any “mandatory governmental licensing, preclearance, or permitting requirement” for AI models: The White House, “Promoting Advanced Artificial Intelligence Innovation and Security”; Skadden analysis. ↩
- 6. That a court already found this administration’s earlier national-security action against Anthropic likely pretextual retaliation: Breaking Defense, “Judge grants Anthropic preliminary injunction” (Judge Lin: the stated reasons “were pretextual” and the “real motive was unlawful retaliation”); Cato Institute. ↩