Meta CEO Proposes Sharing Unreleased AI Models with US Government
Meta CEO Mark Zuckerberg has outlined a new AI policy, proposing that frontier AI labs share intermediate checkpoints of unreleased models with the U.S. government before public launch to enhance security.
Meta Platforms Inc. CEO Mark Zuckerberg has proposed a new approach to artificial intelligence (AI) development and oversight. In an essay published in August 2026, Zuckerberg suggested that leading AI labs should provide the U.S. government with intermediate training checkpoints of models not yet released to the public.
The core of the proposal centers on enhancing security through transparency during the development process. Zuckerberg argues that sharing technical staff and in-progress model checkpoints would enable officials to strengthen critical infrastructure and identify potential security gaps. This, he believes, can be achieved without delaying or restricting public access to new AI products.
Meta rejects the notion that AI safety necessitates concentrating powerful AI within a few trusted entities. The company contends that broadly distributing AI capabilities, including tools for cybersecurity and legal assistance, leads to safer outcomes by preventing any single actor from gaining an excessive advantage. The company cited open-source AI models as an example where widespread deployment helps in identifying and addressing security vulnerabilities more effectively.
As part of its commitment, Meta plans to offer a free tier of its AI tools intended for "billions of people" and resume its releases of open-source models. The company will also implement a new governance structure, granting its independent board the authority to approve safety criteria for model releases and review their compliance.