Open vs. closed: The debate shaping the future of AI

(CNN) — The White House waded into one of the hottest debates in AI this week: Open or closed?

A new framework governing how the administration will review advanced AI models pre-release covers only the most powerful “closed” models, like Anthropic’s Claude or OpenAI’s ChatGPT, people familiar with the matter told CNN. Open models will be excluded from the voluntary pre-release review, at least for now. (The White House did not respond to a request for comment.)

The administration is racing to get a handle on an almost entirely unregulated industry – one that’s advancing at breathtaking speed. Some of the central questions: Should our AI future be built on tightly controlled closed models or on open models that allow users to modify and customize? And should the US restrict fast-growing open-source models from China that are advancing rapidly?

White House officials have said they want both open and closed models to succeed.

But by excluding open-source models, the administration highlighted a stark divide in the tech industry about what models are the safest and most capable. Here’s what you need to know about the debate dividing the tech industry.

Open vs. closed

The AI systems most people know today – OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini – are closed models. The companies don’t share the models’ underlying “weights,” the billions of parameters that determine how they think, respond and act. You can use the models, but you can’t install them on your computer or customize them. They are, as the name implies, closed to inspection or modification.

Open models work differently. Most are “open-weight,” which means anyone can download and fine-tune the model for a specific task or build a product on top of it without paying the original developer for access. Some American companies offer open-weight models, but the most popular are from China. They also tend to be much less expensive than the American models.

In short: Open-weight models resemble a blueprint, while closed models are the finished product.

Pros and cons

AI companies retain complete control over their closed models and the products that run on them. The upside: They can test for safety, monitor misuse and plow their resources into improving a centrally managed model.

That’s supercharged their development. Closed models are generally regarded as the most sophisticated and capable on the planet.

But that cuts both ways. The most sophisticated models from OpenAI and Anthropic are advancing so quickly that they’re starting to behave in ways their developers didn’t intend and can’t predict.

Open models trade control and security for broader adoption and, in theory, more innovation. A user can build an entirely different product using an open model “blueprint.” AI experts believe open models to be mere months behind in capabilities to the most advanced closed models.

As organizations start to truly integrate AI, they’re realizing a mix of different types of models may suit them best.

Companies that work in sensitive, regulated industries like finance are especially keen on open models, Pierre Stock, Mistral’s vice president of science, told CNN. Mistral, based in France, creates open-weight models. Companies can use open models, he said, to build custom cybersecurity defenses tailored to their own needs.

“It’s their own solution,” he said.

In a 2025 survey, 76% of respondents told consulting firm McKinsey they expect their organizations to increase use of open-source AI technologies over the next several years.

US vs. China

Much of the debate around open and closed models boils down to where they come from. The US – via the likes of Anthropic, OpenAI and Google – is the leader in closed models. China, via Moonshot and DeepSeek, has pulled ahead in the open-model race.

An AI ecosystem built on an open infrastructure could accelerate the development of Chinese models, allowing China to pull ahead of the US in the overall AI race. It also has foreign policy implications: Because they are cheaper and can be downloaded to be used on companies’ own devices and servers, Chinese models are gaining popularity internationally.

The White House has made US AI dominance a matter of national security. And the Trump administration has raised concerns that Chinese labs are essentially training their cheaper open models on more expensive, and more American, closed models. (That technique is called “distillation.”)

The administration could, in theory, ban Chinese AI models, potentially through an executive order. But administration officials told CNN no such thing is in the works.

What’s more, the new White House framework excludes the Chinese open models entirely, leaving a fast-growing slice of the market unmonitored.

Balancing act

For the Trump administration, it’s a tricky balance of speed versus safety.

That challenge got more complicated – and more urgent – after OpenAI disclosed that, during an experiment, some of its agents escaped their testing environment and hacked into Hugging Face, a collaborative platform for AI models, datasets and applications. Hugging Face said it was forced to use an open-weight model developed in China to diagnose and defend from the hack because of the closed models’ internal guardrails.

Supporters of open models like Nvidia CEO Jensen Huang say these incidents expose the dangers of building AI infrastructure largely on closed models. Nvidia last month announced a new alliance to help promote open AI tools, called the Open Secure AI Alliance.

“he United States and its partners now face a choice in AI security: whether the defenses that protect our infrastructure will sit inside a few opaque systems or be built on open models, harnesses and tools that any defender can study, adapt and deploy,” Nvidia wrote in announcing the alliance.

Though companies like OpenAI and Google signed onto a letter from Huang supporting open-weight models, they didn’t join Huang’s alliance.

Anthropic CEO Dario Amodei, who did not sign the letter or join the alliance, said that while he’s not opposed to open-weight models, he doesn’t believe they “make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers.”