Who controls access to AI?

Anthropic's latest artificial intelligence model, Fable5, was withdrawn just three days after its launch. Named after the Latin word "fabula," meaning "story," the model arrived with considerable fanfare as the public version of Mythos, an advanced system previously restricted to a small group of users because of concerns about its extraordinary capabilities. Its debut drew global attention. So did its sudden disappearance.

The shutdown was triggered by a single letter from the U.S. government. The letter warned that foreign actors, particularly organizations linked to China, could use Fable5 to obtain sensitive information and called for restrictions on access by foreign nationals. Faced with the prospect of excluding even its own foreign employees, Anthropic ultimately suspended the service.

For many U.S. allies, the episode underscored a troubling reality: access to the world's most advanced AI systems can be cut off overnight if Washington chooses. Increasingly, governments are viewing the issue through the lens of AI sovereignty. Former French Prime Minister Édouard Philippe remarked that the incident showed how infrastructure controlled by others can be disconnected at any moment. The European Commission cited it as a clear example of why Europe must strengthen its technological sovereignty.

As leaders around the world have noted, AI models are rapidly becoming core infrastructure for government administration, defense, education, finance and manufacturing. If access to those systems depends entirely on someone else's permission, serious questions follow. Who decides what constitutes risk? Who determines how AI can be used? Who grants or denies access? The answers increasingly define where power resides.

For countries that have repeatedly witnessed advanced technologies transformed into strategic assets, the Fable5 episode feels familiar. When the United States placed Huawei on its export control list in 2019, companies such as TSMC quickly halted business with the Chinese firm, and even companies in allied nations were required to comply with U.S. rules. The difference this time is that South Korea is no longer merely being asked to align with restrictions imposed on others. It could find itself directly affected by those restrictions. The episode offers three important lessons.

First, dependence on foreign AI systems delivered through cloud-based APIs should be reduced wherever possible. What happens if the foundation model underlying critical AI services suddenly becomes unavailable? That question underscores the need to continue building domestic models and infrastructure for defense, government, security and other strategically important sectors.

Second, countries need the ability to design and operate systems that rely on multiple AI models rather than a single provider or nation. The U.S. Department of Defense reportedly employs a multimodel strategy, assigning different models to different tasks to avoid dependence on any one company. South Korea faces an even greater challenge because it must reduce not only corporate concentration but also dependence on a single country. That calls for a more sophisticated diversification strategy.

Third, countries must secure a position within the AI ecosystem that makes them indispensable partners. As the semiconductor industry demonstrates, the Netherlands holds a seat at every major export-control negotiation because ASML remains the sole supplier of extreme ultraviolet lithography equipment.

The Fable5 episode was not merely a service suspension. It marked the first clear demonstration of who holds the power to grant or deny access to advanced AI. The question facing us now is simple: In the AI ecosystem, will we be among those who can be cut off, or among those who decide who gets cut off?