The White House has started briefing major AI companies on a plan to review frontier models before they ship to the public, according to The Information. The outreach signals the most concrete federal step yet toward pre-deployment oversight of the systems coming out of OpenAI, Anthropic, Google, and others.
What stands out here is the timing. Until now, model evaluations in the US have been voluntary, organized through the AI Safety Institute housed inside NIST. Labs shared early access on a handshake basis. A formal White House review process would change the posture from “please opt in” to “here’s the gate.”
What the briefings cover
The Information reports that administration officials walked AI companies through the framework in private meetings. Specifics on scope, enforcement, and what counts as a “frontier” model weren’t fully detailed in the initial reporting, but the direction is clear:
- Pre-release access to model weights or APIs for government testers
- Evaluations focused on national security risks (cyber, bio, chem, autonomy)
- A structured process replacing the patchwork of voluntary commitments from 2023
Why this matters
Three reasons practitioners should pay attention.
- Release timelines. If a federal review becomes a prerequisite, expect a new step in the launch pipeline for any lab pushing the frontier. That adds weeks, maybe months, depending on how the process is structured. Smaller labs without dedicated policy teams will feel it most.
- The competitive map. Chinese labs aren’t going through US government review. If Washington adds friction on the domestic side without matching export and compute controls, you get the exact “slow our guys down” dynamic that frontier labs have warned about for two years.
- This resets the regulatory baseline. The 2023 voluntary commitments from the Biden era were never binding. A formal review framework, even if narrow, establishes that the federal government has a role in the release decision. That precedent matters more than the first version of the rules.
Context: how we got here
The US has been the outlier among major economies on AI rules. The EU passed the AI Act. The UK set up its own AI Safety Institute and ran pre-deployment tests on GPT-4o and Claude. China requires algorithm registration. Washington stuck with executive orders, voluntary pledges, and NIST guidance.
The AI Safety Institute, created under the Biden executive order, did manage to get testing agreements with OpenAI and Anthropic in 2024. But those were narrow, opt-in, and politically fragile. The current administration’s move suggests it wants something more durable, with clearer authority behind it.
What to watch next
- Which agency runs the review (NIST, Commerce, a new body)
- Whether participation is voluntary or tied to federal contracts and export licenses
- How “frontier” gets defined (compute threshold? capability benchmarks?)
- Industry response from the labs being briefed
The labs have so far played along with voluntary testing because the alternative was worse. A formal framework changes the calculus. Expect lobbying to shape the details before anything gets locked in.
Full details at The Information.