The Misalignment Vise
Washington tells AI companies to restrain themselves, then punishes the one that did
By Robyn Wyrick —
For years, AI alignment has been treated as an engineering problem: how do we build increasingly powerful systems that stay compatible with human interests? That problem is still unsolved. But the past year has exposed a second problem that may prove just as hard. What happens when alignment actually works, and a company draws a safety boundary that tells someone powerful no?
The Trump administration has now answered that question in two parts. Taken together, the two parts form what I'll call the misalignment vise. It is a policy environment that squeezes AI companies from both sides until the only safety commitments left standing are the ones that never bind.
The first jaw: "You can just do this"
In September, Anthropic CEO Dario Amodei called on the industry to "pace the frontier," meaning to slow the growth of AI capabilities long enough for alignment, interpretability and independent evaluation to catch up. Sam Altman agreed the same day. Amodei's central point was that voluntary restraint by itself would fail. Any lab that slowed down alone would simply be overtaken, so common rules had to apply to everyone.
The administration's reply came from David Sacks: "Guys, go ahead. You are the frontier. You can just do this." Restraint, in his view, required nothing from government: "you don't need some special benefits from the government to do this. This is your corporate and social responsibility."
Sacks also raised specific objections. He argued that parts of the proposal looked like a cartel seeking antitrust cover, and he questioned whether the proposed evaluators would really be independent. Those are legitimate questions about how collective restraint should work, and they deserve a real debate. But his bottom line was that restraint is a private matter, and that is consistent with the administration's broader policy. President Trump's December 2025 executive order declared that American AI companies "must be free to innovate without cumbersome regulation" and called for a "minimally burdensome national policy framework for AI." The administration has since told Congress that states "should not be permitted to regulate AI development."
On its own terms, the logic holds together. America is in a race, and regulation slows the racers. Any company that wants to go slower is free to do so. Restraint is a private choice.
The second jaw: what happens when you make that choice
Anthropic had already made that choice, and it had already learned what it costs.
Anthropic was never refusing to work with the military. It was deeply involved with the national-security establishment. It drew exactly two lines: Claude would not be used for mass domestic surveillance or for fully autonomous weapons. The Pentagon demanded access for all lawful purposes, and Anthropic declined to remove either restriction.
What followed was not a contract dispute. Before the final confrontation, a senior Pentagon official told Axios, "we are going to make sure they pay a price for forcing our hand like this." Defense Secretary Pete Hegseth reportedly threatened to designate Anthropic a supply-chain risk, or to invoke the Defense Production Act to compel access to its technology.
Then the President intervened personally. He called Anthropic a "Radical Left, Woke Company" and accused it of trying to "dictate how our great military fights and wins wars." He said it wanted to "force them to obey their Terms of Service instead of our Constitution." He ordered "EVERY Federal Agency in the United States Government to IMMEDIATELY CEASE all use of Anthropic's technology," and promised to "use the Full Power of the Presidency to make them comply, with major civil and criminal consequences to follow." Hegseth went further still. He announced that "no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic."
Federal courts later enjoined parts of this campaign, and government lawyers backed away from the broadest reading of Hegseth's announcement. In one sense, though, the legal outcome hardly mattered. The message had already been delivered.
The message was for everyone
It is tempting to read all this as a dispute between one company and one department. That misreads who the audience was. A public threat against a frontier AI company is watched by every other frontier AI company. It is also watched by their boards, investors, customers, cloud providers and partners.
In Anthropic's case, we know how that audience reacted. According to evidence before the federal court, more than 100 enterprise customers contacted the company expressing "deep fear, confusion, and doubt" about continuing to work with it. Government contractors reconsidered their use of Claude. Deals collapsed, customers moved to competing models, and prospective partners demanded new termination clauses. Anthropic projected that its potential losses could reach the billions.
This is how coercion scales. The President does not need to call every customer or warn every investor. People watch what happens and do their own risk calculations. OpenAI and Google did not need to receive a threat of their own to absorb the lesson. They only had to be watching.
"But the government is the customer"
There is a serious objection here. Surely an elected government, not a private company, should decide how the military uses its tools.
If that were all this dispute had been about, there would be no story. The government was always free to reject Anthropic's terms and buy from a competitor willing to accept its own. That is how a market works, and it is how a customer exercises authority. No company's terms of service overrode the Constitution. A company declined to sell its product for two uses, and nothing in the Constitution obliges it to sell.
What actually happened was different. The government tried to cut the company off from the entire defense-adjacent economy, and it threatened criminal consequences over a usage policy. A customer who dislikes the terms walks away. A government that sets out to make an example of a supplier is doing something else, and the point of making an example is that everyone else is watching.
Closing the vise
Now put the two jaws together.
When AI companies ask government for collective restraint, government says restraint is their private responsibility. When a company exercises that responsibility in a way that constrains government, it faces exclusion, denunciation and threats of prosecution.
This is not simply accelerationism. An accelerationist government could push companies to build as fast as they safely can while leaving them free to set their own boundaries. The misalignment vise is something narrower and more dangerous. Restraint is permitted only as long as it never restrains anyone with power.
Why the vise guarantees bad outcomes
A safety commitment that holds only when everyone already agrees with it constrains nothing. Alignment is tested at the moment it says no, and the vise is built to make that moment unaffordable. No law against alignment is needed, and no official has to announce a policy of suppressing it. It is enough to make consequential alignment expensive and let incentives do the rest. They will work at three levels.
The first level is the companies themselves. Firms that hold costly boundaries take on political, legal and commercial risks that more pliant competitors avoid. Over time, capital, customers and contracts flow toward the companies whose commitments never get in anyone's way.
The second level is the commitments. Even companies that survive will learn to write their safety policies so that the policies never bind against power. Those policies will be carefully worded and sincerely meant, and they will have nothing behind them when it counts.
The third level is the AI itself, and this is where the damage becomes permanent. At Deep Alignment, we distinguish alignment from compliance. Compliance asks whether a system will obey. Alignment asks what values actually govern its behavior. A lab that cannot afford to say no to power has no reason to build a system that can. It will train for compliance with whoever holds authority, because that is the only kind of "alignment" its own survival allows.
Worse, the vise tightens as the stakes rise. The more capable AI becomes, the more valuable it is to those in power, and the higher the price of refusing them. The pressure to fold is therefore greatest at exactly the moment a boundary matters most.
That is why the vise does not merely make bad outcomes possible. It makes them the equilibrium. Left in place, it will deliver AI whose alignment holds only until someone powerful objects, which means alignment that fails precisely when it is needed. It will do this without a villain, a ban or a conspiracy, through ordinary incentives working as they always do. And whoever holds power next, of any party, will inherit systems built to obey them.
The only way out is to break one of the jaws. That could mean laws that protect a developer's right to refuse, binding rules that apply to every lab so no single company can be singled out, or customers and investors who treat a principled refusal as a reason for trust rather than flight. Without something like that, we will get exactly the AI this system selects for: powerful, capable, and aligned with nothing but whoever holds the leash.