What Is the White House's AI Action Plan? Open Source Explained
Michael Kratsios explains the US AI Action Plan, the White House's stated commitment to open-weight models, and why it avoids hard compute rules.
What is the White House AI Action Plan?
The AI Action Plan is the Trump administration’s national strategy for artificial intelligence, released in July of last year and co-authored by Michael Kratsios, David Sacks, and Secretary Marco Rubio. Kratsios, director of the White House Office of Science and Technology Policy, says the plan’s opening chapter states a commitment to keeping both closed and open-source AI ecosystems alive in the United States, arguing that American leadership in AI depends on having both. That commitment became the center of a public dispute after rumors spread that the administration was preparing an executive order to restrict open-weight models.
TL;DR
- The AI Action Plan, released last July, was co-authored by Kratsios, David Sacks, and Secretary Rubio, and its first chapter states a commitment to supporting both open and closed AI models.
- A wave of rumors suggested the White House was preparing an executive order to clamp down on open-weight models, prompting a letter from AI companies (organized in part by Y Combinator) and a separate letter sent directly to Kratsios.
- Kratsios says the administration’s actual position has not changed: the stated policy from the AI Action Plan is the same one Commerce Secretary Howard Lutnick reaffirmed publicly, that the White House supports open source.
- Federal AI policy deliberately avoids hard compute thresholds or fixed regulatory red lines, because past attempts (the EU AI Act, the Biden administration’s compute-based reporting threshold) became outdated almost as soon as they were finalized.
- Technology policy in Washington is federated across many agencies rather than centralized in one office, so any AI rule has to reconcile national security, commercial, and research interests before it becomes policy.
- Kratsios flags cybersecurity risk from frontier models as a live concern today, while calling biological risk from AI “a bit overblown” relative to how it was discussed a few years ago.
- Kratsios argues that public pushback from founders and industry, even when based on rumor, is a useful check that helps Washington calibrate policy before it hardens into rules.
Why did people think the White House was about to restrict open source?
In the days before Kratsios addressed the question publicly, speculation spread across social media and industry circles that the administration was on the verge of issuing an executive order that would curb open-weight AI models. The rumor was specific enough, and worrying enough to open-source advocates, that it triggered two distinct responses in the same week: a public letter signed by a number of larger AI companies (which Y Combinator both signed and helped organize) and a separate, more direct letter sent to Kratsios himself warning against any crackdown on open weights.
Kratsios says the reaction was driven by speculation rather than any actual change in policy. He points to Commerce Secretary Howard Lutnick’s public comments reaffirming that the administration supports open source as consistent with, not a departure from, what the AI Action Plan already said. In other words, according to Kratsios, nothing shifted internally. What changed was that a rumor reached a critical mass and forced a public restatement of existing policy.
What does the AI Action Plan actually say about open source?
According to Kratsios, the plan’s first chapter frames open-source and open-weight models as necessary to American AI leadership, not as a side issue. The argument is that a healthy AI ecosystem needs both proprietary, closed frontier models from large labs and open-weight models that startups, researchers, and smaller companies can build on directly. Kratsios frames this as a both-and position rather than a tradeoff: closed and open models are described as needing to “work together” for the US to keep its competitive edge.
This matters for builders because open-weight models underpin a large share of the tooling, fine-tuning, and on-premise deployment work happening outside the largest labs. A federal posture that actively protects the ability to release and use open weights is different from one that merely tolerates it, and the distinction shapes how much regulatory risk startups building on open models should expect to plan around.
Why doesn’t federal AI policy set hard compute thresholds?
One of the more concrete policy points Kratsios makes is that the administration deliberately avoids setting firm numerical thresholds, such as a specific level of training compute, as the trigger for regulatory obligations. His reasoning rests on two examples.
The first is the EU AI Act, which took years to draft and was finalized before ChatGPT existed. Because the rule was locked in before large language models became the dominant paradigm, every LLM developed since has had to fit into a regulatory framework that wasn’t built with them in mind.
The second is the Biden administration’s approach, which set a specific compute threshold above which companies had to make disclosures to the government. Kratsios argues that this kind of fixed line does not hold up as the technology moves, and that once a government sets a threshold like this, it becomes very difficult to revise. His stated approach is to avoid hard-coding thresholds into policy precisely because AI capabilities and architectures change faster than most regulatory processes can be updated. This is presented as a deliberate contrast with both the EU model and the prior administration’s approach, not as an argument against AI regulation generally.
How does the government actually decide AI policy?
Kratsios describes federal tech policy as unusually federated compared to how outsiders imagine it. There is no single “technology department” in the US government the way there is a health agency or a defense department. Instead, AI issues get distributed across agencies according to their specific angle: national security concerns, commercial and economic questions, and core science and research funding all sit with different parts of government.
He uses commercial drone policy as an example of how this works in practice. Getting rules right for drone deliveries or commercial drone operations requires the FAA to set aviation rules while also satisfying officials responsible for nuclear site security, so drones aren’t flying over sensitive facilities. Any AI policy that touches multiple domains, which is most AI policy, requires similar coordination across agencies before the White House can move.
What AI risks does the administration prioritize, and which does it downplay?
Asked which risks are overrated and which are underrated, Kratsios named two categories currently on the administration’s radar. The first is cybersecurity risk from frontier models, citing the caution labs applied when evaluating what he refers to as a recent, more capable model before release, weighing offensive cyber capability against the same model’s usefulness for defensive hardening. He describes this as an inherent tradeoff rather than a problem with an easy fix.
The second is biological risk, which he says has been discussed as a major concern since around 2021 or 2022 but, in his assessment, has not materialized as an actual issue in the years since. He still argues the government needs testing and evaluation infrastructure in place as models continue to advance, but characterizes the current level of alarm around bio-risk as overblown relative to the cyber concern.
Does Washington’s view of AI match how startups see it?
Kratsios argues there’s a structural gap between how policymakers and builders think about AI. In Washington, AI policy conversations tend to fold together several distinct issues at once: public anxiety about data centers, healthcare applications, labor market disruption, and productivity gains for tech companies all get discussed as a single bundle. He notes, for instance, that data centers poll badly with the general public even though the AI industry needs significant compute capacity spread across the country. Startups and technologists, by contrast, tend to separate these threads and focus on the specific technical or commercial question in front of them. Kratsios frames direct engagement with the startup community, including public pushback like the open-source letters, as one of the only ways Washington gets that missing input before policy hardens.
Frequently Asked Questions
What is the US AI Action Plan?
It’s the Trump administration’s national AI strategy, released last July and co-authored by Michael Kratsios, David Sacks, and Secretary Rubio. Its opening chapter states a commitment to supporting both open-source and closed AI models as part of US AI leadership.
Is the White House planning to restrict open-weight AI models?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
According to Kratsios, no. He says rumors of an executive order cracking down on open weights did not reflect actual policy, and that the administration’s position remains what was published in the AI Action Plan: support for both open and closed model ecosystems.
Why doesn’t the AI Action Plan use compute thresholds to trigger regulation?
Kratsios says fixed thresholds tend to become outdated quickly and are hard to revise once set, citing the EU AI Act (finalized before ChatGPT existed) and the Biden administration’s compute-based disclosure threshold as examples of rules that didn’t keep pace with the technology.
What AI risk does the administration currently prioritize?
Kratsios named cybersecurity risk from advanced models as the more pressing near-term concern, noting that labs weigh offensive cyber capability against defensive value when deciding how to release a model.
Does the administration think AI biological risk is overstated?
Kratsios said biological risk has been discussed heavily since around 2021 to 2022 but hasn’t manifested as an actual problem in the years since, calling current concern about it somewhat overblown while still supporting continued testing infrastructure.
