the Obernolte-Trahan draft invites two discussions: 'would this be good for frontier safety' and 'is that worth preemption and political tradeoffs'. the latter is important and perhaps decisive, but I think it's worth stating plainly: if passed, this would be good for AI safety.

my initial read is that, fundamentally, the draft federalises a stronger version of the state law of the land contained in CA's SB-53, NY's RAISE, and IL's SB-315. but starting with the minimal case: even if the draft only elevated the existing state-level provisions to the federal level and did nothing else at all, there would be a lot to like about that. SB-315, the Illinois bill dealing with indepdent verification organisations, is much more effective on the federal level because one of the main bottlenecks for IVOs right now is a durable demand signal - a federal bill does much more to fix that than a state bill.

there's also a real precedent-setting path dependency here: no matter whether this draft becomes law or not, we will probably do more frontier safety legislation in '27. that is also the argument for a federal version of SB-53 and RAISE: it creates the enforcement capacity on the federal level instead, in time to build on it with further legislation down the road. no matter how effective the Californian and New York enforcement are, they do not build directly into the eventually-necessary federal capactiy for enforcement that we will eventually clearly want.

to contextualise that point, my main worry for the coming federal AI policy conversations in the next Congress is not so much that we will lack the political capital to make it happen; it's that we will scramble for solutions without sufficient federal capacity to enact them. remedying that is high on my list, and lifting the impressive wins safety advocates achieved on the state level to the federal level is an obvious way to do that. I value this more highly than the prospect of slightly better enforcement over the next year or two.

on top of the federalisation, the draft also throws in a lot of plainly extremely helpful additions. the first is $100m and codification for CAISI. this is a meaningful budget increase and much-needed anchoring that would be particularly helpful in an environment where CAISI has been sidelined by the recent EO. for good reason, this has long been a central safety ask. days after the EO has pushed some central evaluation capacity into the IC instead, that's particularly welcome.

the second is a broader pipeline to put the IVO mechanism to work, starting with injunctive relief through state AGs against developers that don't allow third-party audits - ensuring the verification and audits actually do happen, and not even concentrating the power to enforce and ensure this evaluation only with the Trump administration. in combination, those two come as close to building an actual architecture for regulating frontier development as we've seen in US legislation: CAISI can guide and hold the IVOs to some minimal substantive standards, the IVOs can assess whether the labs comply, and the AGs can ensure the IVOs actually get to evaluate. this combination is a big deal, and well beyond what people expected this kind of deal to include.

i sympathise with everyone who argues that that is not enough, but I don't think it needs to be right now. a few months back, I argued we could muddle through via three steps: pass transparency bills, build up a powerful IVO environment, and then give that ecosystem binding power to enforce helpful safety standards. this draft clears the first two bars in a way that clearly sets up a direct path toward the third. i find it hard to argue that the trajectory of frontier AI safety wouldn't be safer for it.

one argument to the contrary would be: the deal for safety advocates keeps getting better; we started with empty preemption a year ago, now we're here, why not wait another year until the deal is even better as a result of rising salience and more state-level wins? i think two trade-offs pull against that view. the first is time - we're quickly getting to more and more powerful systems, and I'd rather get the federal infrastructure in place faster than slower. the second is the influx of politics - increased salience also means other types of politicians and politics crowd in, more visible issues become more important, and the relative influence of the safety interests wanes. on safety alone, this bill seems good enough to me to make the jump to the federal level right now.

zooming back out to the politics, what should safety advocates make of that? there are real substantive and tactical reasons not to bite the bullet on the broader preemption that the draft entails - especially for a bill that is not likely to pass on the merits. the preemption is real and broad, and there are both policy and political reasons to oppose that fiercly, especially within the Democratic party: there are undoubtedly bills dealing with other harms caused by AI systems that would not pass if this draft became law. that is a very real cost, and it's reason enough to oppose it or to call it a bad draft in total - but not to call it bad for frontier safety on the merits.

so of course, you should not expect every safety organisation to have the luxury of assessing this draft on these narrow safety merits alone. but there's still a real tension here: a lot of safety advocacy trades on the premise that safety concerns are of urgent and singular importance, and that time is of the essence. it seems difficult to justify that the default safetyist reaction to a draft that makes as much real progress is to plainly reject it out of hand in grounds that only tangentially relate to the actual safety merits. my - perhaps delusional - hope is that the public conversation can reflect this: safety advocates don't have to carry this to victory, but I think it would still be unfortunate for the safety movement to become one of the sharpest critics of a draft that would make frontier AI safer.

so amid the ad-hoc responses you'll read on this draft, I think it's very much worth taking the step back and evaluating it on its safety merits alone. i think the result of that is clear: if you specifically care about AI safety, there's much to like about the actual substance of this draft.