U.S. Pushes Meta to Open Its AI Models for Government Review

U.S. Pushes Meta to Open Its AI Models for Government Review

The Trump administration is pressing Meta to join a voluntary federal review program for advanced artificial intelligence models, making the Facebook parent one of the most visible holdouts among major American AI developers.

According to a New York Times report cited by Reuters, OpenAI, Anthropic, Google DeepMind, Microsoft and xAI have already agreed to give the U.S. government early access to their most advanced AI models for national-security testing. Meta has not yet finalized an agreement, though the company has said it is working through the details and hopes to sign soon.

The request is officially voluntary. But it comes at a time when Washington is becoming more aggressive about frontier AI, especially systems that could help bad actors carry out cyberattacks, military misuse, or other high-risk activity.

What Washington Wants From Meta

The review program is being handled through the Commerce Department’s Center for AI Standards and Innovation, or CAISI. The office is the renamed successor to the U.S. AI Safety Institute and is becoming one of Washington’s main tools for evaluating powerful AI systems before they are widely released.

Under an executive order signed by President Donald Trump in June, leading AI developers are asked to provide the government with early access to their most capable “frontier” models before public launch. The review period can last up to 30 days.

During that time, government experts can test what the models are capable of, where they may fail, and whether they could help malicious users perform dangerous tasks, such as finding software vulnerabilities or supporting cyberattacks.

The order says the process is voluntary. It does not create a licensing system, a pre-clearance rule, or a requirement that companies must obtain government permission before releasing an AI model.

That distinction matters. This is not full government control of AI releases. But it is still a major shift. When nearly every major U.S. AI company agrees to the same review process, the company that has not yet joined stands out. Right now, that company is Meta.

Why Meta May Be Hesitating

Meta’s hesitation is not surprising.

The company has spent years promoting itself as a supporter of open AI. Unlike some rivals that keep their most powerful systems behind closed products, Meta has released open models that developers, researchers and startups can download and build on.

That strategy has helped Meta position itself as a company pushing AI beyond closed labs and into the wider technology community. But allowing public downloads of some models is not the same thing as giving the federal government early access to inspect future frontier systems before release.

For Meta, the concern may be about precedent. A voluntary review system today could become a stronger expectation tomorrow. Even if the government is not asking for formal approval rights, companies may worry that refusing a review could create political or regulatory pressure.

There is also the competitive angle. AI development is now a global race involving American companies, Chinese firms and governments around the world. A delay of even a few weeks could matter if a rival launches first, captures developer attention, or becomes the new benchmark for the industry.

At the same time, Meta cannot easily ignore Washington’s request. The company is already under intense scrutiny over privacy, competition and platform safety. Standing apart from other major AI labs on national-security reviews could invite more questions from regulators and lawmakers.

The Security Question

The debate comes down to a difficult question: how much access should the government have to the most powerful AI systems before the public does?

Supporters of the review program argue that the government needs early visibility into frontier models because the risks are no longer theoretical. Advanced AI systems can help write code, search for vulnerabilities, generate convincing text, and automate complex tasks. In the wrong hands, those capabilities could make cyberattacks faster and more effective.

That is why CAISI is focusing heavily on security testing. The goal is to understand whether a model could meaningfully assist a malicious actor before the model becomes widely available.

The White House order also attempts to address concerns from AI companies. It says access should be handled with protections for confidentiality, cybersecurity, intellectual property and nondisclosure. In simple terms, the government is trying to say: let us test the risks, but we will protect your trade secrets.

Still, the concern remains. For an AI company, its model behavior, architecture, safety techniques and training process can be among its most valuable assets. Even controlled government access can feel risky if a company fears leaks, insider threats or competitive exposure.

The Shadow Over “Voluntary”

Although the review process is voluntary, recent events show why companies may feel pressure to cooperate.

Earlier this month, Reuters reported that the U.S. government ordered Anthropic to suspend access to its most advanced models, Fable 5 and Mythos 5, for foreign nationals because of national-security concerns. Anthropic later said it would disable access to those models for all users after receiving the order.

That episode showed that Washington is willing to act directly when officials believe a powerful AI model creates a security risk.

So while the CAISI review program is not mandatory, it exists in a political environment where the government is already taking a stronger role in AI oversight. A voluntary request from a regulator with enforcement power can feel very different from a casual invitation.

This does not mean the U.S. has created government control over AI releases. It has not. But the boundary between voluntary cooperation and expected compliance is becoming less clear.

A New Normal for Big Tech

If Meta signs the agreement, every major American AI developer would be part of a review system that did not exist in this form a year ago. That would mark a major change in how frontier AI is handled in the United States.

For supporters, this is a practical compromise. It gives the government a chance to test powerful systems before launch while avoiding a heavy licensing regime that could slow innovation.

For critics, it may not go far enough. A system based on company cooperation can look more like an honor code than true regulation. If participation is optional, then the final decisions still sit largely with private companies.

But even a voluntary program can become powerful if it turns into an industry norm. Once most major companies participate, refusing to join may create reputational, political and business pressure.

For Meta, the choice is bigger than one agreement. Signing could reduce pressure from Washington and show that the company is aligned with national-security expectations. Refusing could strengthen Meta’s image as an independent champion of open AI, but it could also make the company look isolated at a sensitive moment.

What Comes Next

Meta’s next move will decide whether the U.S. review program becomes a near-industry standard for frontier AI.

The program remains voluntary, and the White House order says it does not create a licensing or pre-clearance requirement. But Washington is clearly seeking earlier visibility into the most powerful AI systems before they are released.

Most major American AI developers have reportedly agreed to participate. Meta is now under pressure to decide whether it will join them.

If Meta signs, the U.S. will move closer to an industry-wide review norm for advanced AI models. If it does not, the debate over voluntary oversight, government pressure and Big Tech independence will likely grow louder.

Either way, the direction is clear: the most powerful AI models are no longer being treated only as commercial products. They are increasingly being viewed as national-security assets.