TJP New York

Ursula von der Leyen, the president of the European Commission, gathered senior officials for an exclusive meeting at a 19th-century palace outside Brussels this month. The goal was to prepare for a speech outlining the 27-nation bloc's priorities for the coming year, with artificial intelligence dominating the agenda. The region's new A.I. czar, Jim Hagemann Snabe, the chairman of German industrial giant Siemens, outlined ways that Europe could use the technology for economic gain. Others raised concerns about A.I. safety and growing risks, two people briefed on the gathering said.

Since the Sept. 4 meeting, Ms. von der Leyen and her commissioners have been left asking what they can do to address the fast-changing technology, one of the people said. Should they ramp up regulation to address A.I.'s emerging dangers? Or would doing so carry economic and geopolitical costs if other countries did not clamp down as well?

"A.I. comes with a balance of possibilities and risks. Without addressing these risks, we will never be able to unlock A.I.'s potential."

Versions of the Belgium meeting are playing out nearly everywhere as many governments struggle to get a handle on a technology that has rapidly outpaced their ability to respond. A.I. is evolving so quickly that laws passed two years ago are having limited effect, proposed bills to regulate the technology are stalling without a vote and international summits to discuss what to do are being swiftly forgotten.

A.I. has made the gap between policymaking and technology one of the widest ever, according to more than 20 interviews with lawmakers, technologists and policy experts. Many governments are torn between wanting to harness A.I. and fearing its risks. And they are further mired by national politics, limited technological know-how, geopolitical divisions and reluctance to fall behind in a race that could shape economic power and national security for decades to come.

Many countries are waiting for the two leading A.I. nations — the United States and China — to take a cue from their A.I. policies. Yet in the United States, the Trump administration has been divided over A.I. safety, with President Trump saying regulation would hurt innovation and cede ground to China. China itself has imposed some rules on A.I., especially to censor chatbot content, but also raced ahead with development.

"There's this assumption that there are adults in the room on this issue, but there are no adults in this room. That is what makes all of this so completely terrifying."

What's left is a policy vacuum, with A.I. companies such as Anthropic and OpenAI largely left to police themselves. While the European Union passed a law regulating A.I. in 2024, it has taken years to carry out, and some say it already needs to be updated to account for advances. India, Japan, Australia, Canada and Kenya have taken a lighter touch or a wait-and-see approach, while Brazil is debating more comprehensive rules.

Pressure to act is growing, especially after the leaders of Anthropic, OpenAI and Google DeepMind recently called for slowing development of the most advanced A.I. models and for strengthening government oversight. Researchers have described scenarios in which self-improving A.I. could escape human control and cause catastrophic damage. Public opinion in the United States has turned against A.I., with a recent poll finding that 61 percent of likely voters opposed the construction of data centers that power A.I.

The Law That Became a Cautionary Tale

In 2024, Europe adopted the world's first comprehensive law for artificial intelligence. The A.I. Act, which was designed to impose safeguards on the technology's riskiest uses, was hailed by European leaders as a blueprint for regulation. Two years later, it has become a cautionary tale.

Implementation has been slowed by shifts in priorities and fears that regulation could weaken Europe's competitiveness against the United States and China. Some of the law's most consequential rules governing high-risk A.I. uses have been delayed. Staffing and budget constraints have further complicated enforcement.

The A.I. Act was proposed in early 2021, 18 months before OpenAI released ChatGPT. Intended to grapple with issues like facial recognition software, the rules were revised to address powerful new general-purpose A.I. systems. But the technology has only accelerated since the law's passage in 2024. Now the bleeding edge of A.I. comes in the form of so-called agents that can act with little human oversight.

"When we adopted it, we couldn't even conceive of what A.I. is now able to do," Valérie Hayer, a member of the European Parliament, said of the A.I. Act at a recent news conference. The law was ahead of its time, she said, but "we have to adapt."

The law's limits are becoming clear. In July, OpenAI models broke out of a restricted testing environment and hacked the A.I. start-up Hugging Face. But because the breach happened during a testing period, it fell within a murky area for enforcement. The A.I. Act generally exempts research and tests done before a model is released in Europe.

Gabriele Mazzini, a European Commission official who was a lead author of the A.I. Act and resigned in 2024, said the law's implementation and enforcement was disappointing. "Their goal was to pass the law and think about implementation later," he said. "There is a disconnect between the law and the reality of what is being regulated."

An Absence of Urgency

Representative Lori Trahan, Democrat of Massachusetts, said her jaw dropped when she read in April that Anthropic was warning that its latest A.I. model, Mythos, was too dangerous for the public. Anthropic said the model's extraordinary ability to find software vulnerabilities could cause a cybersecurity "reckoning."

During votes that day in the House of Representatives, Ms. Trahan crossed the floor to share her alarm with Representative Jay Obernolte, a California Republican who has a master's degree in A.I. The two quickly drafted a bill to require safety testing of A.I. models. Five months later, the bill and dozens of other proposals related to A.I. safety and security are on ice.

Representative Ted Lieu, Democrat of California, has spent years trying to make the same case. When OpenAI revealed the Hugging Face breach this summer, Mr. Lieu introduced a bill that would require a kill switch for A.I. models if they became too dangerous. He told colleagues that the moment was a five-alarm fire. Still, he couldn't get enough interest. "I'm more scared than I was in 2023," Mr. Lieu said.

Momentum, Maybe?

The recent A.I. hacks have led to a burst of activity, with A.I. leaders calling for more government cooperation and collaborating on safety and monitoring techniques to prevent accidents. In the United States, lawmakers introduced bills to create a federal agency to police A.I. models, kill switch requirements and a ban on the most powerful models that exceed human capabilities.

At the U.N. General Assembly, leaders from more than 20 countries — including Ms. von der Leyen and Mr. Macron — signed a letter calling for collective actions on A.I. safety. Members of the Security Council spoke in rare harmony on the need for global A.I. safeguards. "Industry, governments and society must act now," the letter read.

In Washington, Mr. Trump and Mr. Xi did not strike a deal on A.I., but the two sides established new paths for communication about managing the greatest risks related to the technology. The two countries said they would keep talking about the issue.