"AI Will Kill Us"

AI: synthetic image.

Written in August, rewritten on 27 October 2026.

For some months now, the men who build artificial intelligence have spent part of their time explaining that it might destroy us. Bill Gates gave it a dozen pages in August. Elon Musk was calling for a moratorium back in 2023. Dario Amodei, who runs Anthropic, forecasts the disappearance of half of all junior office jobs. In the autumn they signed a joint call to slow down.

I read those texts in full. There is a pattern in them.

The race

Gates dismisses slowing down in two lines, halfway through his essay: if anyone had a credible plan to slow AI worldwide he would probably support it, but he does not think it will happen. The inevitable is set down as a fact, without being demonstrated. Everything that follows manages the damage: companies take the gains, states pay for retraining.

Musk signed the March 2023 letter calling for a six-month pause in AI development. Four weeks later he set up his own artificial intelligence company. And that pause, had it been granted, would have frozen the leader for six months — exactly the time a latecomer needs to get into position.

Anthropic, for its part, proposes real rules, and that deserves credit: limits it publishes and undertakes not to cross, support for the Californian law the rest of the industry was fighting. They are rules Anthropic already complies with.

All three alarms have the same shape: not one threatens the position of the person raising it. The joint call in the autumn is no exception. We slow down, provided everyone slows down. Nobody brakes alone. And I do not see the Chinese, the Russians or any of the other lunatics slowing down on their side.

What if the fear is justified?

On social media the answer arrives as a picture: a robot, a red eye, a machine killing everything that moves. That is the weakest scenario of them all. There are four plausible arguments.

A person uses the machine to build a virus: the only short path, and the only one where a handful of people can reach the whole world. Cyberattacks on power grids, water, hospitals: many deaths, not humanity. The machine advising the military: a war started by mistake — in 1983 the Soviet officer Stanislav Petrov refused to believe the screen telling him five American missiles were inbound, and he was right. And finally the machine that escapes its master. That is the argument everyone talks about, and the weakest: you only have to cut the power, and that is the end of the story.

What holds the first one back? A trained virologist does not need the machine, he already knows. The one it helps is the determined amateur — and the determined amateur still has to culture it, handle it and protect himself, in a laboratory you do not improvise in a garage. Above all he needs the raw material. Synthetic DNA is ordered by post these days; reputable suppliers check every order against a list of dangerous sequences, and in most countries nothing obliges them to. That is where the rules should go. Not on the models.

The virus that would carry us all off needs, on top of that, a combination never yet seen. Ebola kills one patient in two and has never set off an epidemic outside Africa: a patient flat on his back infects nobody. Covid killed fewer than one in a hundred of those it infected, and went round the world three times. Highly contagious and highly lethal at the same time does not exist in nature.

One thing is rarely told. In the 1990s the Aum Shinrikyo sect had the money, qualified scientists, laboratories and the intent. It failed: in 1993 it released anthrax over a Tokyo neighbourhood, and nobody fell ill. It fell back on sarin in the underground.

We never waited for anyone

Humanity has never needed a machine to kill itself in large numbers. Last year the Uppsala researchers counted 65 conflicts involving a state: more than ever since they began counting, which is to say since 1946.

Ukraine, in 2025 alone (94,700). The First World War (15 to 22 million). The Second (70 to 85 million), the Holocaust among them (6 million). Stalin (6 to 9 million). Mao’s Great Leap Forward (35 to 45 million). The Khmer Rouge (1.5 to 2 million). Rwanda, in a hundred days (500,000 to 1 million).

The people in the field have been asked. Nearly 2,800 researchers answered in 2023: between 38% and 51% of them give at least a one-in-ten chance to an outcome as bad as human extinction. That is not a forecast. It is a polite way of saying they do not know.

And whatever origin one ascribes to Covid, the most disruptive pandemic in a century needed no artificial intelligence at all.

Today I am likelier to be killed by my neighbour than by a lunatic AI.

Sources