It codifies the Center for AI Standards and Innovation
Miles Brundage#20
Here's a joint opinion piece from @JayObernolte and @RepLoriTrahan on their draft framework: "AI holds an extraordinary promise, but the promise alone is not a strategy."
https://news.bloomberglaw.com/legal-exchange-insights-and-commentary/america-needs-one-national-framework-for-artificial-intelligence
Positive users praised the Obernolte-Trahan federal AI framework draft as a smart collaborative foundation and victory, while negative users condemned its preemption of state AI laws as an industry handout that sells out constituents.
The Trahan-Obernolte discussion draft text has dropped.
I'm going to publish a longer and more carefully considered piece on this in the near future, but quick takes for twitter:
I think the substance of the catastrophic risk provisions (IVO-style auditing and SB 53-style transparency, both with fairly robust rulemaking authority; incident reporting; whistleblower protections) is quite good, for the most part. Not absolutely perfect, but better than any other federal cat-risk bill that's been introduced so far.
Unfortunately, I also think that the preemption (all state laws regulating development) is far too broad, and that this makes the bill as drafted a bad deal overall. Currently, you're preempting (for example) a broad range of state child safety laws, in exchange for a federal framework that (from what I've seen so far; I have not yet read all 270 pages) doesn't include substantial federal child safety protections. And it's not just child safety, it's every possible current or future state law that regulates development.
If you narrowed the preemption to avoid that kind of asymmetry between the policies you're implementing federally and the policies you're eliminating at the state level, I think I would strongly support a bill like this. It doesn't have to be completely 1:1, necessarily; something like @deanwball's proposal, which preempted five specific categories of state law in exchange for federal transparency, might be good enough (depending on the details).
Tl;dr. -- the draft is quite good on substance, but quite bad on preemption, and therefore (IMO) a bad deal overall. My hope is that it's a step in the right direction, and that as discussion continues the preemption section can evolve into something that makes more sense in terms of both politics and policy.
The Obernolte-Trahan draft is the best federal AI proposal released yet. It contains provisions for AI lab transparency, lab employee whistleblower protections, a framework for independent verification orgs, funding for CAISI and public compute (NAIRR), and much more.
The bill largely focuses on the development of AI systems, so it preempts state governments from passing conflicting regulations in that domain while preserving their ability to regulate deployments. This is similar to how vehicle safety standards come from the federal government, but states regulate driver licensing, vehicle registration, auto sales (via regulation of car dealerships), road safety (via traffic law), and adjudication of post-incident harms (via traffic court, tort liability for drivers, etc.).
You will notice that the role of states in the car example is not small! Yet vehicle safety standards, because they affect a thing that is mass manufactured at great upfront capital expense, need to be federal. The deployment/development division of labor is the most straightforward way to do AI preemption out there, and the time for federal preemption has long since come.
This is the best AI law draft America has yet produced, and I applaud Representatives Obernolte and Trahan for their tremendous effort and leadership. I wish everyone in Congress held themselves to their standard.
The threats AI poses to our national security, safety and workforce are here and growing by the day.
@JayObernolte and I released draft legislation to protect workers, establish real accountability for the most powerful AI systems, and position the U.S. to lead on AI.
Reps @JayObernolte and @RepLoriTrahan have released a discussion draft of a federal framework for AI. This is a serious, substantive effort to govern how the most powerful AI systems in the world are built, tested, and deployed – and putting it out takes real political courage. Credit to both of them, full stop. A few immediate thoughts on the draft itself, from a first read:
It's a strong bill
Pretty much agree with Andrew here, which shouldn’t be a surprise as I pretty much agree with Andrew on most things.
Reps @JayObernolte and @RepLoriTrahan have released a discussion draft of a federal framework for AI. This is a serious, substantive effort to govern how the most powerful AI systems in the world are built, tested, and deployed – and putting it out takes real political courage. Credit to both of them, full stop. A few immediate thoughts on the draft itself, from a first read:
I think there are plenty of child safety laws that regulate development.
Imagine a child safety law designed to prevent things like the recent xAI scandal where Grok was being used to digitally undress underage girls. Suppose this law says "AI developers are required to ensure that their models have guardrails that prevent the creation of NCII involving minors."
IMO, that law is (a) a child safety law, that (b) regulates AI development. Therefore, it would be preempted by the Trahan-Obernolte bill. Does that make sense? Agree/disagree?
Live reactions to this:
- defines catastrophic risk as death/serious injury to >50ppl or more than $1b in damage; calls out CBRN and loss of control risks
- critical safety incident defined as modification or exfiltration of weights, failure of risk mitigation measures, or loss of control
- frontier model defined as trained on >10^26 FLOPs
- frontier developer defined as trained a frontier model AND has >$50m in revenue
This discussion draft is the product of bipartisan conversations, including with our colleagues @RepSuhas, @RepScottPeters, @RepFranklin & @RepHouchin.
We hope to build on this work with input from workers, stakeholders and members of the public. Join us! https://trahan.house.gov/news/documentsingle.aspx?DocumentID=3783
- formally authorizes CAISI
- 🚨 tasks it with administering an independent verification organization regime, including by licensing IVOs
- gives it $100m annual budget
A quick note up top: releasing as a discussion draft was a smart way to invite others into the conversation. It is a strong foundation built to be pressure-tested and improved in the open — in their words, “the start of a serious national conversation.” On an issue this consequential, we all have a responsibility to engage and to do so in good faith.
@CharlieBull0ck why do you think it preempts a broad range of child safety laws? aren't those counted as laws concerned with deployment rather than development?
That’s a novel formulation for preemption — as should be the case, for such a complex and quickly-moving technology that requires regulatory nimbleness. But to get this right, the community needs to stress-test and better define the specific real-world implications of this language construction for states. We would not be in favor of preemption if it does not meaningfully and defensibly retain states rights to protect the civil rights of its citizens broadly, or the safety of its children specifically.
@RepLoriTrahan @JayObernolte Pre-emption 🧐
A lot here is genuinely very good. First: the draft establishes ongoing independent verification for catastrophic risks. If you know Fathom, you know IVOs. We continue to strongly believe that catalyzing a robust marketplace of third-party verification for AI products is the single-most important step that can be taken to ensure AI is safe, secure, and worthy of earned trust.
This draft sets up mandatory independent verification – only for catastrophic risk — by expert organizations licensed by CAISI to provide ongoing oversight of frontier models. If a system is out of compliance or poses an imminent risk, the IVO notifies the US Attorney General and the AGs of state that opt in, who can then seek injunctive relief in court or pursue fines. There are outstanding questions regarding the scope, authority, and responsibilities of these IVOs, and we look forward to working with the authors to address them. Still, that’s real accountability — a mechanism to stop systems that pose serious risks — with due process.
Importantly: these IVOs will not simply be responsible for checking that the frontier companies are doing what they say they are doing. IVOs are required to independently check the models themselves and, using their own tools and methods, verify that they do not pose a threat of catastrophic harm. This is not a mere transparency audit.
We continue to believe that for non-catastrophic risks, the IVO system should be voluntary. That has been the focus of our work in the states. Sufficient incentives exist to impel developers to participate: market differentiation and competitive advantage, legal clarity (i.e. standard of care defined upfront by experts), and access to insurance. But for catastrophic risk, given the gravity and potential to inflict massive American casualties, we can see why the co-authors chose to make this framework mandatory at the federal level.
On this provision and others, we’ll continue to engage and do our part to make this bill the strongest, best effort at making the most consequential technology of our time safe, secure, and beneficial to Americans. This is our chance to get it right.
The hardest question is preemption. The draft draws its line at model development (one federal standard for how frontier models are built and made safe), while attempting to explicitly preserve states’ power to protect their own residents from AI that’s deployed in unsafe ways.
The draft contains many other strong provisions: meaningful transparency requirements, safety and governance framework requirements, and protections for whistleblowers who flag wrongdoing. That these critical elements are table stakes for a federal framework is a credit to the work of several states that have fought hard battles to bring transparency into the system. It is also quite notable that – in the context of these hard-fought state policies – the authors include a genuine enforcement role for state AGs should the federal government fail to act, and meaningful power for states to help enforce the transparency and verification provisions.
It codifies the Center for AI Standards and Innovation
Miles Brundage#20Here's a joint opinion piece from @JayObernolte and @RepLoriTrahan on their draft framework: "AI holds an extraordinary promise, but the promise alone is not a strategy."
https://news.bloomberglaw.com/legal-exchange-insights-and-commentary/america-needs-one-national-framework-for-artificial-intelligence