Washington Moves to Test Advanced AI Models Before They Reach the Public

By Tahani Elghazaly
Published May 6, 2026

Latest news on WhatsApp أجدد الأخبار على واتساب

The United States is entering a new phase in its approach to advanced artificial intelligence, expanding a government program that reviews powerful AI models before they are released to the public.

The Center for AI Standards and Innovation, known as CAISI, operates within the National Institute of Standards and Technology at the U.S. Department of Commerce. It has announced new agreements with Google DeepMind, Microsoft and xAI, allowing government experts to conduct pre-deployment evaluations, post-deployment assessments and targeted research on frontier AI systems.

The move builds on earlier voluntary work with OpenAI and Anthropic, which had already been cooperating with U.S. government scientists to test unreleased models and identify vulnerabilities.

The agreements do not mean Washington is now issuing a formal launch license for every AI model. But they do mark a more organized and earlier government role in reviewing the most advanced systems before they reach wide public use.

According to Reuters, U.S. government scientists are focused on demonstrable risks, including whether advanced models could be used to launch cyberattacks on American infrastructure, help adversaries develop chemical or biological weapons, or corrupt the data used to train U.S. AI systems.

The significance of the move lies in a bigger question: who reviews AI before millions of people can use it? Until now, technology companies have carried much of the responsibility for internal testing, with some voluntary outside reviews. Washington is now trying to secure an earlier seat at the table before the most sensitive models are released.

NIST says CAISI has already completed more than 40 evaluations, including on state-of-the-art models that remain unreleased. The new agreements also allow evaluators from across government to participate in assessments, including testing in classified environments when needed.

One important detail is that developers may provide models with safeguards reduced or removed so evaluators can better understand their true capabilities and risks. This kind of testing is designed to simulate how malicious actors might try to misuse a model.

For everyday users, the story may sound technical. But it touches the future of anyone using AI for work, study, programming, writing, health, law or research. As models become more powerful, the question is no longer only whether they can produce accurate answers. It is also whether they can be misused at scale.

The issue now is not whether AI will continue to advance. It is how governments and companies can reduce the chances that the most powerful models become tools for hackers, hostile actors or dangerous misuse before society understands their full capabilities.

For major technology companies, the message is clear: fast innovation is no longer enough. Trust, testing and proof of safety may become part of the competition itself.