
Hi
Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you:
Why the safest US AI strategy is a national one
Building institutions that can keep up with AI
If you find them helpful, make sure you’re signed up for the next issue.
[Policy] No patchwork, just leadership
Our view on AI safety has been straightforward for months: the best way to govern the most powerful AI systems is through clear national standards that are prevention-first, innovation-friendly, and built to address the highest-consequence risks before harm occurs.
That case is becoming more urgent by the day. China’s updated, government-approved 5-year plan calls for integrating AI across most sectors of its economy, not just the tech industry, underscoring its goal to lead on AI. In that context, how the United States governs AI is not just a domestic policy question – it’s a question of whether we can move with comparable clarity, speed, and coherence.
That’s why we’re encouraged to see growing momentum toward a stronger federal framework in the US.
The basic logic is simple: AI leadership is increasingly an issue of national security, and national security requires prevention rather than punishment after the fact.
When states act alone, they are often pushed toward liability-based approaches, holding companies accountable only after something has already gone wrong. But preventing the most serious risks requires technical expertise, threat intelligence, and access to classified systems that only the federal government possesses.
We already have a strong building block for that approach: the Center for AI Standards and Innovation (CAISI). CAISI gives the federal government the ability to test and evaluate advanced AI systems before they are widely deployed. This is exactly the kind of centralized capacity needed to manage risks that no single state can address on its own.
We are also starting to see alignment across levels of government. Recently passed AI legislation in both California and New York has begun to incorporate CAISI-aligned standards and evaluation frameworks, pointing toward a more coherent national model rather than a fragmented patchwork. That kind of harmonization, where states build on federal technical standards rather than diverging from them, is a promising signal for what a durable system could look like.
In that context, we continue to believe a clear order of approaches makes the most sense:
First, the US should establish a federal framework through legislation that requires frontier labs to test their systems using classified government capabilities. This would allow the government to assess model capabilities, set national safety standards, and address both catastrophic risks and issues like protecting children. Any preemption would be drawn to help ensure the usefulness of the national safety standards. This approach would provide strategic coherence, prioritize prevention, and position the US to help shape global AI safety standards grounded in democratic principles.
Second, in the absence of a national framework, states should align around the emerging model in California and New York. With nearly 2,000 AI-related bills proposed at the state level, consistency matters. A common, harmonized approach would allow the US to continue leading on innovation while establishing de facto national standards, even if those approaches rely more on liability after harm rather than prevention before it.
Third, if states continue to move forward, they should include mechanisms that encourage federal leadership. California and New York have begun to do this by recognizing that meaningful federal testing processes could satisfy state-level requirements. This creates an incentive for a national framework, while acknowledging that only the federal government can move from a liability model to a prevention model.
We’ve done this before. During the Space Race, a unified national strategy helped put humans on the Moon. In aviation, a national approach to safety became the global standard. AI presents a similar test. Done right, the US AI strategy can both strengthen safety at home and help set the terms for how AI is governed around the world. – Chris Lehane, Chief Global Affairs Officer, OpenAI

[News] Institutions for AI that won’t stand still
For decades, the technologies that defined national security (from radar to the early internet) were built inside the government. AI breaks that pattern. Today, the most consequential systems are being developed in the private sector, evolving faster than traditional policy frameworks were designed to handle.
We believe that this shift is already reshaping how governments and companies work together, and that the next phase of AI governance will hinge less on static rules and more on building credible, adaptive institutions.
A central focus: the growing role of AI Safety Institutes (AISIs). These independent bodies are emerging as critical partners in evaluating frontier models, testing safeguards, and advancing the science of AI risk assessment. OpenAI has worked closely with groups like the UK AI Security Institute and the U.S. Center for AI Standards and Innovation (CAISI). But our call goes further: no single institution can fully understand increasingly powerful systems. What’s needed is a global network – so sharing evaluation methods, coordinating research, and enabling continuous, independent analysis as models evolve.
That global layer will matter, but so will national capacity. In the US, CAISI is a potential cornerstone institution, one that can help policymakers assess frontier capabilities, compare systems developed by strategic competitors, and evaluate performance in sensitive or classified environments.
The bottom line: in the Intelligence Age, governance isn’t just about rules; it’s about building the institutions that make those rules credible, durable, and fit for technologies that won’t stand still. – Sasha Baker, Head of National Security Policy, OpenAI

[About] OpenAI Forum
Explore Forum programming by and for our community of 60,000 AI experts and enthusiasts from across tech, science, medicine, education, government, and other fields.
5:00 PM – 6:00 PM EST on Mar 18

[About] OpenAI Academy
The Academy is OpenAI’s free online and in-person AI literacy trainings for beginners through experts.
OpenAI has called for a nationwide AI education strategy – rooted in local communities in partnership with American companies – to help our current workforce and students become AI-ready, bolster the economy, and secure America’s continued leadership on innovation.
1:00 PM – 2:00 PM EST on Mar 19
[Disclosure]
Graphics created by Base Three using ChatGPT.





I'm glad to see more detail on what you'd support in a federal framework; thanks for writing this!
I'm confused by this part, however:
> "With nearly 2,000 AI-related bills proposed at the state level, consistency matters."
What is the data-source you're using for what counts as an "AI-related bill"?
When I've investigated this in the past, the state bill count has been substantially overstated. Many are not really AI-related; have no regulatory impact; or are even favorable to AI's rollout.
Commonly people would claim ">1,000" and the ones that matter for frontier AI would be more like a few dozen at most
You can find my analysis at https://stevenadler.substack.com/p/mythbusting-the-supposed-1000-ai