The die is cast.

Human beings have always told stories about knowledge that cannot be taken back. Pandora opens the jar. Prometheus brings fire to humanity and pays for it. The bag of adverse winds is opened and the entire journey of Odysseus has to start again.

These stories endure because all major scientific and technological discoveries change the world. No good deed goes unpunished because no deed goes unpunished; everything has consequences.

Everything.

Once people learn to make something powerful, a warning about its dangers cannot make them forget the knowledge. Artificial intelligence belongs in that long tradition. Its creators can choose what to build next and how carefully to deploy it. They cannot return us to a world in which the technology is unknown.

Former Google CEO Eric Schmidt put the practical problem plainly this week: “We’re not going to do a pause.” Asked about a mutual agreement between the United States and China, he came back to two obstacles: incentives and verification. “How do you verify?”

He is right, of course.

That is the AI race in the language of the prisoner’s dilemma. Both countries might prefer a slower, safer pace. But each fears that the other would keep advancing behind closed doors. If Washington pauses while Beijing continues, America could surrender a consequential advantage. Beijing has every reason to ask the same question about Washington.

Neither, needless to say, will pause. The more experts warn about the dangers of this new tech, the less likely each nation will be to pause anything.

The problem exists within countries, too. A company that stops training new models has to trust its competitors to do likewise. Inspectors could monitor the construction of enormous data centers and track some shipments of advanced chips. They would have a much harder time proving that no laboratory was improving an existing model, experimenting privately or finding a more efficient way to train one.

President Donald Trump sees the strategic stakes. 

Speaking at the United Nations on Tuesday, he said, “Whoever wins AI” wins, and promised to encourage its development while watching it closely. He also rejected an international scheme to control the technology. His newly preferred name, “super intelligence,” may or may not catch on. His central question will: What happens if America deliberately slows while a competitor does not?

China is hardly acting like a country prepared to wait. 

This week Alibaba announced plans for a much larger AI model, unveiled a new chip and set an ambitious data-center expansion target. The future model and chip shipments are plans, not achievements already in hand. But the direction is unmistakable. Stanford’s 2026 AI Index found that American and Chinese models had traded the lead and that the measured gap between their best models was small as of March. America’s broader AI ecosystem remains stronger; its lead in model performance is less secure.

None of this means the warnings should be brushed aside, of course.

At Wednesday’s UN Security Council meeting, Anthropic CEO Dario Amodei warned that poorly managed AI could pose a risk to humanity. OpenAI CEO Sam Altman argued that decisions about its future require democratic processes and accountable governments. They are asking serious questions about systems that are growing more capable and, in some cases, able to act with less direct human supervision.

In fact, modern AI has repeatedly surprised the people building it. Researchers can measure impressive performance on difficult tasks while still finding strange failures on simpler ones. Stanford calls this a “jagged frontier”: a model can excel at competition mathematics and stumble over an ordinary visual task. That unevenness argues for testing, scrutiny and humility before powerful systems are entrusted with consequential work.

It does not, by itself, answer Schmidt’s verification question. 

Declaring knowledge forbidden is easier than keeping it forbidden. Nor does the impossibility of a universal pause mean governments must accept every use of AI. They can require serious incidents to be reported, test high-risk systems before deployment, limit an agent’s access to sensitive networks and insist on human approval before consequential actions. Companies can be held responsible when negligent safeguards cause harm.

Schmidt himself favors agreements designed to prevent “surprises” between the United States and China. An emergency notification rule would not require either country to stop developing AI. It would give both a way to explain a dangerous incident before an accident is mistaken for an attack. That is a more achievable form of cooperation because neither side benefits from stumbling into a crisis.

Pandora’s story is often told as a warning about opening the jar. 

Unfortunately (or fortunately), we are past that decision. The die is cast.

The knowledge exists, the models are spreading, and competitors are pressing ahead. Our decision now is whether to meet the risks with rules we can actually enforce — and whether America will remain capable enough to help shape the world those rules must govern.

(Contributing writer, Brooke Bell)