Elsewhere I took issue with the pro-moratorium advocates’ strategic approach – not on the basis of the policy itself (of which I am not in favor) – but purely on the basis that it did not account for any dynamic political response, resulting in suboptimal outcome.
In other words, it predictably backfired.
But one element I did not address, is with the stated policy goal itself which has now reared its head in the most obnoxious way in the form of the Fable export control intervention.
To catch up: the Department of Commerce insisted Fable be disabled for all foreign users based on a national security risk, which effectively required Anthropic to disable access for all users.
I didn’t predict this exactly, but directionally it's not surprising. Especially with this Administration. I did note:
Moratoire ≠ Laissez Le Bon Temps Roule
I characterize the pro-moratorium model as thus:
The premise: advancements in AI are moving so fast, here and around the world, and innovation, if allowed to continue apace, can deliver incredible benefits for humanity.
The problem: states are rushing ahead with thousands of misguided laws and regulations that impose unnecessary costs without commensurate benefits, further compounded by inconsistencies and complexities across the 50 jurisdictions. It’s impossible for innovators to maximally do good work while navigating this policy thicket.
The solution: Congress should pass a law that prevents unproductive non-federal laws via Article I’s preemption authority to maintain a favorable national policy environment, for up to 10 years.
This is a simplified but I believe fair narrative characterization.1 I’ve omitted from the solution their call for an affirmative federal governance law. I’ve run through in the same earlier piece why an actual substantive policy framework was never legible, and therefore undermined its credibility. So it doesn’t need to be reckoned with here. I’ll add to earlier arguments though, there’s no structural reason an actual federal framework couldn’t pass alongside - in the same leg text - the moratorium language (of course obviating the need for the latter, so…).
Taking this premise/problem/solution structure at face value, the solution was never going to result in what they actually wanted, at least not fully. New laws, federal, state or otherwise, yes change the status quo and very often interfere with otherwise effective market institutions. Yet even poorly designed laws, when stable, can be adapted to by market participants in nearly optimal ways such that further changes should be very carefully considered because they are likely to make an imperfect situation worse.
But an absence of laws does not mean an absence of policy. As I noted in a testimony, using an overwrought metaphor: policy energy will move but rarely destroyed.
You Didn’t, So I Will
Administrations have always exercised their regulatory authority, increasingly in the last few, to apply old laws to new circumstances. This Admin in particular has been, let’s say, extremely creative, in squeezing new policy toothpaste from old tubes. This is a White House that apparently seriously considered widespread suspension of habeas corpus. Consider tariffs under IEEPA, removal orders under Alien Enemies Act, and others, some much more narrow in scope. And those are ones asserting substantively new authorities from extant law. You could multiply the list several-fold if you include potentially defensible authorities effected through categorical neglect of required procedural requirements (DoD’s designation of Anthropic as supply chain risk, also Commerce’s export control order for the same, fall into this category).
It didn’t start in 2025. Under Biden, some of the COVID actions (CDC’s eviction moratorium, OSHA’s vaccine mandate) and student loan forgiveness under HEROES Act, all resemble the same Executive application of laws beyond plausible authorities.
If you assume the current Executive is favorable to you, it’s not unreasonable to assume this isn’t a risk. But it’s always a non-zero risk. No one can say, “we had no way to know how mercurial this President is” – it’s a second term!
There’s no way to know the specific path of AI development, and therefore how different implications will assume salience or irrelevance. But again no reasonable person should credibly claim it will be predictably linear and therefore a straight extrapolation from the current moment. In fact the premise of both utopians and doomers is that this is a transformative technology, not repetitive to prior ones.
An affirmative law, or several which together make an effectively comprehensive federal governance framework, can mitigate much of the Knightian uncertainty of the policy environment. It won’t eliminate Executives from eisegetical statutory interpretation. But it can fill a negative space and simultaneously suppress prior law which almost certainly is less suitable.
In the absence of such a suited-for-purpose federal statutory framework, we have today’s latest contretemps. Even if Anthropic is able to seek remedy through court or negotiate a favorable resolution with the Admin - and I’ll be sympathetic to anyone arguing otherwise - this is a decidedly worse environment than most any plausible suboptimal Congressional deal.
Even with an affirmative statutory framework defining the sidelines of the policy space, could an impetuous executive still take the same action? Sure. But again it would reduce the likelihood and strengthen the argument for courts to overturn (which is admittedly not costless in any case).
We Find for the Plaintiff
The President and hyperactive Cabinet Secretaries are not the only risk. An emerging cohort of lawsuits are using existing law to assert harms. Under fairly plain vanilla state consumer protection laws and the like, leading labs may be eventually subject to really onerous precedent.
If the plaintiffs are successful, don’t assume it’ll be any easier to modify those laws in a more favorable way ex post. An already unpopular industry, having been characterized as the “new tobacco” by crusading state politicians or entrepreneurial trial lawyers will see to that. Constituencies will not find much sympathy for eye-popping jury awards, even if there is a compelling argument they are misguided.
Even if plaintiffs are unsuccessful, months or years of pernicious discovery will serve as an interim win. There’s no version of a case which gets past summary judgement that doesn’t disadvantage the reputation and possibly actual operations of the defendant. One does not have to jump to conspiratorial speculation about “smoking guns” or the fabled magic memo proving unscrupulous behavior. Good faith and free-flowing communication between employees will always reveal something exploitable as leverage in a future fight.
Nature Abhors a (Policy) Vacuum
Dean Ball characterizes the current policy node, the one reflecting the Fable 5 action, as effectively existential to the industry. He lays some contributory blame on Anthropic’s behavior, which he argues was avoidable. But he was down this road going back to the Pentagon action.2
So with the benefit of hindsight, if we accept Dean’s characterization, we ended up at a much worse non-stationary equilibrium, whereas we could have moved towards a suboptimal but better one with more stability with a broadly applicable sensible federal law. Almost any law that was not limited to “10 year moratorium.”
I say benefit of hindsight. Fully concede where we are today was not foreseeable ex ante. But that’s the point – to reduce the decision space, reduce the uncertainty, and try to drive toward a stable policy environment that, hey, may even be Pareto improving. And even industry responses to unproductive rules can be welfare-enhancing.3
I must concede there is not one organizational advocate but hundreds of companies, scholars, thought leaders, and so on I am reducing to a single synecdoche. In particular I lean heavily on a16z and connected coalitions because they are both rhetorically voluminous, oft-quoted, and represent the largest till of explicitly political funding on this issue. That said, many smart people could argue their own aligned position is nevertheless distinguishable. Against which I wouldn’t argue, but simplification is necessary to make the point.
https://www.aeaweb.org/articles?id=10.1257/pol.20190019