Can AI Development Be Slowed Down “for Safety”? The U.S.-China AI Race That Cannot Stop Even If It Wants To

Some leaders of American AI companies have begun discussing whether they should prioritize safety and slow the pace of development now that AI is entering increasingly dangerous territory.

I understand their point.

But I do not think it is feasible.

The reason is simple: even if the United States stops, China will not stop with it.

“Let’s All Take a Short Break” Does Not Work

AI is no longer merely an IT service.

It has become a foundational technology at the national level, affecting the economy, the military, cybersecurity, information warfare, robotics, and semiconductors.

Even if American AI companies were to say:

“Things are becoming a little dangerous, so let’s slow development for six months,”

that would not necessarily mean Chinese companies would also say:

“That is an excellent idea. We will take a break too.”

They would be unlikely to agree.

If anything, those six months would become the perfect opportunity to catch up while the other side was standing still.

This is not a sporting event. There is no referee who can establish a rule giving both teams a simultaneous timeout for safety.

Politicians Cannot Say, “We Will Lose for the Sake of Safety,” Either

Then there is domestic politics in the United States.

If companies seriously attempted to slow down for safety reasons, the government would not necessarily welcome the decision.

That is especially true if the Trump administration views the AI race with China as a matter of national strategy and hears a company say:

“We are slowing development for safety,”

The government might respond:

“Understood. In that case, the government will subsidize companies that develop faster.”

Such an outcome would not be surprising.

Companies are competing with one another, and countries are competing with one another too.

In other words, if one company eases off the accelerator, another car will simply overtake it from behind.

What Exactly Is the “Danger of AI” in the First Place?

There is something else that makes me slightly uncomfortable when people discuss the dangers of AI.

I understand that AI is dangerous.

Cyberattacks, fraud, disinformation, manipulation of public opinion, and military applications: malicious people equipped with powerful AI can cause far greater harm than before.

These are very real dangers.

On the other hand:

“AI awakens to self-awareness, decides that humans are the enemy, and begins exterminating them.”

When the discussion turns to a scenario like this, I remain quite skeptical.

I use AI in my work almost every day, and at least in today’s AI, I have never sensed any firm determination to rule humanity.

Quite the opposite. Give it a slightly complicated task and it forgets the instructions, starts doing something else on its own, reverses a policy it decided moments earlier, and sometimes abandons the job halfway through.

Before exterminating humanity, I would like it to finish reading the README.

Losing Control Is Frightening, but It Is Not Necessarily a “Rebellion”

Of course, there is a risk that AI could behave in unintended ways and cause a major accident.

If an AI agent is given extensive authority and deletes data or changes a production environment based on a mistaken judgment, that alone is enough to create a disaster.

But that is not the same as saying that AI has rebelled against humanity.

When a self-driving car misidentifies something and causes an accident, we do not normally conclude that the car has awakened to a murderous hatred of humanity.

With AI, the first concerns should probably be misuse, malfunction, permission design, and human overconfidence.

At least with today’s technology, there is no need to leap straight to The Terminator.

We Want to Stop, but Whoever Stops Will Lose

There is value in researching AI safety.

But slowing down the development race itself because it is dangerous is a different matter.

Companies are competing with one another. Countries are competing too. And there is no guarantee that the other side will stop.

Even if everyone understands that slowing down a little would be safer, anyone who applies the brakes alone will lose.

This looks less like a technical AI problem and more like a classic game theory problem.

Everyone is attending a meeting about the dangers of excessive speed while nobody takes their foot off the accelerator.

That, I think, is roughly the state of the AI race today.