AI Almost Started A War With China - And We Nearly Didn't Catch It

For years, the nightmare scenario surrounding artificial intelligence and warfare has looked something like science fiction: autonomous weapons making their own decisions, computers selecting targets, machines eventually removing human beings from the process of deciding who lives and who dies.

But we may have been worrying about the wrong nightmare.

AI does not need control of a missile to help start a war. It only needs to convince a human being that the missile needs to be launched.

A remarkable incident inside the U.S. military earlier this year offers a glimpse of just how dangerous that possibility has become. According to CNN, an intelligence report circulated within the military claiming that a Chinese vessel traveling in the Middle East was carrying components connected to a nuclear weapons program.

The report set off alarms.

American forces began preparing to intercept the vessel. Armed personnel were reportedly preparing to board it. Military aircraft were sent into the air.

Then, just before the operation moved forward, officials took a closer look.

The intelligence was wrong.

Not slightly wrong. According to one source familiar with the incident, it was "entirely false." Investigators discovered that an analyst had used artificial intelligence in producing the report and that a chatbot had incorrectly identified what the Chinese ship was carrying.

One source told CNN the incident "almost started a war."

That may sound dramatic. But consider what could have happened if American military personnel had actually attempted to board a Chinese vessel based upon a false accusation involving nuclear weapons.

This was not an AI laboratory experiment.

Military forces were reportedly already moving.

One of the most common reassurances about military artificial intelligence is that humans will remain in control.

A machine might analyze intelligence. It might identify threats. It might recommend targets. It might help commanders make decisions.

But ultimately, we are told, a human being will make the final call.

This incident exposes the weakness in that reassurance.

Humans can remain completely in control while still making disastrous decisions based upon information supplied by machines.

The AI doesn't need authority to order an attack. It can simply provide the intelligence upon which a human commander orders one.

That distinction becomes increasingly important as the Pentagon moves aggressively to integrate AI throughout the military. Its 2026 AI strategy calls for accelerated integration across warfighting, intelligence and enterprise operations, including AI-enabled battle management and decision support stretching from campaign planning into the "kill chain." The Defense Department has been pursuing AI for command and control, operational planning, intelligence activities, autonomous systems and other military applications for years.

The objective is understandable: speed.

Modern warfare moves extraordinarily quickly. The military that identifies a threat, understands it and responds first may gain an enormous advantage.

But speed comes with a terrifying downside.

It gives you less time to discover that you are wrong.

That may be the most disturbing question emerging from this episode.

Someone apparently did what we would hope a human being would do: questioned the intelligence.

Officials dug deeper.

The error was discovered.

The operation did not proceed.

But what happens next time?

What happens when the intelligence appears even more convincing? What happens during an actual crisis when commanders believe they have minutes rather than hours to react? What happens when several AI systems independently repeat information that originated from the same false source?

Most importantly, what happens when the opposing military begins reacting before anyone discovers the mistake?

Imagine a future confrontation in the Pacific.

An AI system mistakenly identifies preparations for a Chinese attack. American commanders respond by repositioning aircraft and ships. Chinese surveillance detects those movements and interprets them as preparations for an American strike.

China responds.

American AI systems detect China's response and conclude that the original warning was correct.

More forces move.

Missiles are placed on alert.

Commanders suddenly have minutes to determine whether the enemy is preparing to attack.

Nobody wanted war.

But each side's response makes the other side increasingly convinced that war is coming.

That is how an AI hallucination could become a military reality.

The danger becomes even greater because America cannot develop military AI in isolation.

China is pursuing many of the same advantages.

Both militaries understand the strategic value of processing enormous quantities of information faster than human analysts ever could. Both have powerful incentives to use AI to shorten the time between detecting something and responding to it.

That creates an extraordinary technological arms race.

If one country can identify targets in ten minutes, the other wants to do it in five.

Then someone wants to do it in sixty seconds.

Eventually the technological goal becomes almost instantaneous detection, analysis and response.

But there is a paradox buried inside that achievement.

The faster machines help nations make decisions, the less time humans have to discover that the machines made a mistake.

We already know how dangerous false warnings can be.

During the Cold War, both the United States and Soviet Union experienced incidents in which warning systems produced information that could have been interpreted as evidence of an attack. Human skepticism, additional verification and simple time helped prevent erroneous information from becoming catastrophe.

AI could make that safety margin increasingly thin.

That is why the most realistic danger of military AI may look nothing like the movies.

There doesn't need to be a robot rebellion.

There doesn't need to be a conscious machine.

There doesn't even need to be an autonomous weapon.

There only needs to be bad information delivered at precisely the wrong moment and trusted by precisely the right person.

And today's generative AI systems have an uncomfortable characteristic: they can be confidently wrong.

Anyone who has extensively used AI has probably encountered it. Ask a chatbot about an obscure subject and occasionally it will produce a remarkably detailed answer containing information that simply isn't true.

In ordinary life, that might mean a fake quotation, nonexistent source or incorrect historical fact.

Inside an intelligence system, the consequences are entirely different.

A hallucinated restaurant recommendation is annoying.

A hallucinated nuclear-weapons shipment can move military forces.

None of this means militaries will abandon artificial intelligence.

Quite the opposite.

The Pentagon's current strategy explicitly seeks accelerated military AI integration, while China is pursuing AI-assisted military command and decision-making of its own. Neither side wants to discover during a future confrontation that its opponent can process the battlefield faster.

That makes this incident particularly important.

We have spent years debating what could happen someday when AI becomes powerful enough to threaten humanity.

Meanwhile, a much simpler danger may already be here.

AI doesn't have to become smarter than mankind. It doesn't have to become conscious. It doesn't have to seize control of nuclear weapons.

It only has to be wrong.

Wrong about a ship.

Wrong about a missile.

Wrong about an aircraft.

Wrong about what another country intends to do.

And eventually it has to be wrong at a moment when there isn't enough time for someone to catch the mistake.

This time, someone checked.

This time, the aircraft returned without an international confrontation.

This time, the Chinese vessel wasn't boarded.

But the military race toward faster AI-assisted decision-making is accelerating, not slowing down.

For generations, one of the greatest fears of the nuclear age was that the wrong person might someday put a finger on the button.

The AI age presents a different danger.

The human finger may remain firmly on the button.

The machine may simply tell him when to push it.