There is a draft AI bill (discussed here, Politico reports it's being pitched by Anthropic) that has been making the rounds. I have some takes that I think differ somewhat from those I've been seeing so far in the AI-safety-adjacent discourse, centered around interpretation of specific portions that I think may be under-appreciated in how people are reading the bill. I walk through those here.
TL;DR
An explicit statutory path to block a model would probably make a federal block more likely, relative to current law
The draft includes provisions that seem to preserve significant federal authority to "block" AI labs (Section 12(a)(2))
The preemption section, as written, doesn’t preempt state suspension of models. If that is "fixed" other parts of the preemption provision would probably preserve a lot of state law AI claims, like public nuisance suits.
I'm primarily interested in questions like those raised here about whether this draft bill makes it harder for courts to block or restrict AI labs. The draft bill refers to a "blocking order" as "an order suspending the development, deployment, or internal use of a frontier model on the ground that the model presents an imminent enumerated risk". I think that is a good baseline for what it means for a court to "block" the actions of an AI lab. I have in mind here a slightly broader definition, something like "an order suspending development, deployment, or internal use".
I'm only considering the draft bill vs current federal law, not a comparison to something like the FRONTIER Act.
Nothing I say here is meant to suggest the draft bill is overall good or bad. I just think some of these details are being glossed over.
Details people may have missed
Statutory path to blocking
Courts can't block the actions of AI labs on their own initiative; they need a case in which that is an appropriate remedy. It isn't clear to me that there is a direct route to a federal cause of action for an AI system that poses a catastrophic risk under current law. There are definitely arguments that one could make under various laws, but courts may be hesitant to read existing law too broadly. A big impact I think a bill like this would have is that it would create specific statutory authority, approved by Congress, for a court to do some blocking. I think this is actually the predominant effect the bill would have on the likelihood of a federal block, and as a result I think it would increase, rather than decrease, the likelihood that a court would block actions by an AI lab (again, relative to status-quo).
The draft does impose some important limitations, such as the action needing to come from the Commerce Secretary (so a presidential administration that doesn't want to act is a big impediment), putting the burden on the Secretary, requiring de novo review at various stages, requiring a showing that a blocking order is the least restrictive option that addresses the risk, expiring a blocking order after 45 days, and barring consideration of evidence a lab hasn't seen.
Reasonable people can disagree on how the balance of these factors pans out. My read of the recent history of US federal courts is that standing is becoming a very relevant concern and plaintiffs with creative theories may face a more hostile environment, making an explicit statutory mechanism highly valuable. That inclination informs my take that the explicit federal route would be a pretty big deal.
Section 12(a)(2)
Section 12 deals with orders restricting the actions of AI labs. Section 12(a)(2) specifically disclaims effects on other laws:
No department, agency, or officer of the United States may suspend the development, deployment, or internal use of a frontier model under this Act except pursuant to an order entered or issued under this section. Nothing in this section affects authority under any provision of law other than this Act.
I think this would mean that the bill wouldn't substantially decrease the likelihood of a court blocking the actions of an AI lab under other federal statutes. This would also seem to cover executive agency action that doesn't need involvement of the courts as well.
Section 14(b)(1)
Section 14(b) is the preemption provision. As written, this includes Section 14(a)(1)-(3) in what is preempted, leaving out Section 14(a)(4) "suspension of models". This would seem to preempt parts of state laws like NY's RAISE Act or California's SB 53 to the extent they impose disclosure, evaluation, or incident-reporting obligations, but also allow state law activity on "suspension of models". It's unclear to me whether this is because there is meant to be a Section 14(b)(2) that addresses how preemption as applied to Section 14(a)(4) is meant to work, or whether (4) was left out of Section 14(b)(1) accidentally.
Section 14(a)(4) also includes the limitation that "This paragraph shall have no effect unless section 12 is enacted". Section 12 is the "blocking orders" section. This suggests to me that the drafters' intention is that preemption of "suspension of models" would only occur if there is also a federal route for this (under Section 12). That is very different from some prior advocacy of federal preemption of AI legislation (e.g. the 2025 Senate reconciliation bill "moratorium"), where the desire seems to be to lock-in a regime of relatively limited regulation. I think Section 14(d) and (e) and Section 15(c) also support this conclusion. Section 15(c) includes "It is the intent of Congress that State authority be displaced under section 14 only so long as the Federal requirements of this Act remain in force", which seems to suggest pretty strongly the goal here isn't to preempt for purposes of blocking regulation in general.
If we assume Section 14(b)(1) is updated to include all of Section 14(a)(1)-(4), I think Section 14(c)-(g) still preserve a lot of state law claims, such as public nuisance suits against AI labs. Section 14(a)(4) does include "judicial decree", while Section 14(c)(5) preserves state authority to apply its common law, so it isn't definitive. Arguably Section 14(e) would mean that ties go to the state.
Worth watching
Updated drafts may clarify how preemption is meant to work with regard to Section 14(a)(4).
Suncor (argued Oct. 5 at SCOTUS, awaiting decision) addresses preemption in the context of public nuisance suits against energy companies for their conduct relating to climate change. This is the route that Section 14 potentially leaves open.
Whether Section 12(a)(2) gets weakened or strengthened, and if any information comes out about how it's meant to function within the context of the draft bill.
There is a draft AI bill (discussed here, Politico reports it's being pitched by Anthropic) that has been making the rounds. I have some takes that I think differ somewhat from those I've been seeing so far in the AI-safety-adjacent discourse, centered around interpretation of specific portions that I think may be under-appreciated in how people are reading the bill. I walk through those here.
TL;DR
Scope
Details people may have missed
Statutory path to blocking
Courts can't block the actions of AI labs on their own initiative; they need a case in which that is an appropriate remedy. It isn't clear to me that there is a direct route to a federal cause of action for an AI system that poses a catastrophic risk under current law. There are definitely arguments that one could make under various laws, but courts may be hesitant to read existing law too broadly. A big impact I think a bill like this would have is that it would create specific statutory authority, approved by Congress, for a court to do some blocking. I think this is actually the predominant effect the bill would have on the likelihood of a federal block, and as a result I think it would increase, rather than decrease, the likelihood that a court would block actions by an AI lab (again, relative to status-quo).
The draft does impose some important limitations, such as the action needing to come from the Commerce Secretary (so a presidential administration that doesn't want to act is a big impediment), putting the burden on the Secretary, requiring de novo review at various stages, requiring a showing that a blocking order is the least restrictive option that addresses the risk, expiring a blocking order after 45 days, and barring consideration of evidence a lab hasn't seen.
Reasonable people can disagree on how the balance of these factors pans out. My read of the recent history of US federal courts is that standing is becoming a very relevant concern and plaintiffs with creative theories may face a more hostile environment, making an explicit statutory mechanism highly valuable. That inclination informs my take that the explicit federal route would be a pretty big deal.
Section 12(a)(2)
Section 12 deals with orders restricting the actions of AI labs. Section 12(a)(2) specifically disclaims effects on other laws:
I think this would mean that the bill wouldn't substantially decrease the likelihood of a court blocking the actions of an AI lab under other federal statutes. This would also seem to cover executive agency action that doesn't need involvement of the courts as well.
Section 14(b)(1)
Section 14(b) is the preemption provision. As written, this includes Section 14(a)(1)-(3) in what is preempted, leaving out Section 14(a)(4) "suspension of models". This would seem to preempt parts of state laws like NY's RAISE Act or California's SB 53 to the extent they impose disclosure, evaluation, or incident-reporting obligations, but also allow state law activity on "suspension of models". It's unclear to me whether this is because there is meant to be a Section 14(b)(2) that addresses how preemption as applied to Section 14(a)(4) is meant to work, or whether (4) was left out of Section 14(b)(1) accidentally.
Section 14(a)(4) also includes the limitation that "This paragraph shall have no effect unless section 12 is enacted". Section 12 is the "blocking orders" section. This suggests to me that the drafters' intention is that preemption of "suspension of models" would only occur if there is also a federal route for this (under Section 12). That is very different from some prior advocacy of federal preemption of AI legislation (e.g. the 2025 Senate reconciliation bill "moratorium"), where the desire seems to be to lock-in a regime of relatively limited regulation. I think Section 14(d) and (e) and Section 15(c) also support this conclusion. Section 15(c) includes "It is the intent of Congress that State authority be displaced under section 14 only so long as the Federal requirements of this Act remain in force", which seems to suggest pretty strongly the goal here isn't to preempt for purposes of blocking regulation in general.
If we assume Section 14(b)(1) is updated to include all of Section 14(a)(1)-(4), I think Section 14(c)-(g) still preserve a lot of state law claims, such as public nuisance suits against AI labs. Section 14(a)(4) does include "judicial decree", while Section 14(c)(5) preserves state authority to apply its common law, so it isn't definitive. Arguably Section 14(e) would mean that ties go to the state.
Worth watching