Google, Microsoft and xAI to let the US vet AI models before release

Google DeepMind, Microsoft and xAI have agreed to let the US government review new AI models before their public release. That means pre-release government review is moving from an experimental program to a mandatory stage of development for the largest companies.

How the review works

The Center for AI Standards and Innovation (CAISI), part of the US Department of Commerce, runs pre-release assessments of new AI models well ahead of their public launch. It analyzes how systems perform on coding, math and science tasks, identifies potential risks, and works with companies to improve safety before a model ships.

The program began operating in 2024 with OpenAI and Anthropic. In its first year and a half, it has completed 40 reviews. Both companies recently renegotiated their partnerships to align with the Trump administration's new priorities.

Google DeepMind, Microsoft and xAI are joining on new terms. Practically all the key developers of frontier AI models now coordinate their releases with the government.

Why the government oversees AI

The US administration has set CAISI several goals: understand what frontier AI systems are actually capable of; identify potential risks before models reach the mass market; and control exports of AI technology in the interest of national security.

CAISI conducts targeted research alongside the companies, testing models for release readiness. This includes analyzing how a system handles potentially dangerous scenarios and checking what new capabilities have appeared compared with the previous generation.

The administration's priorities for AI development:

• Guarantees that systems remain controllable
• Full understanding of frontier model capabilities
• Transparency toward regulators
• Protection of national security
• Control over technology exports

What this means for the industry

The largest companies' agreement to pre-release government review means that releasing new frontier models in the US without coordinating with CAISI is no longer possible. For companies, this slows the development cycle by several weeks before release. For the government, it provides leverage over the direction of development. For users, it potentially means higher safety standards for the models they use every day.

What it means

The era of complete freedom for big AI companies in deciding when to release models is coming to an end. The state is taking a more active role in the governance of the largest technologies, which affect millions of people. This may speed up the development of safety standards, but slow the pace of innovation.