He Won’t Call It a World War. I Will.

I watched one of the most powerful men in technology get worried on a laptop screen in a cave in rural Spain, the laptop running off the same solar panels that keep the lights on and the kettle honest. It was a YouTube video, the sort that arrives uninvited in your recommendations with a title in capital letters and a thumbnail of a serious man looking into the middle distance. I almost scrolled past. I’m glad I didn’t, although “glad” is doing some heavy lifting in that sentence.

The serious man was Eric Schmidt, and the reason I stayed was a lesson I learned long before YouTube existed. In the RAF Police, and in the years of security management that followed, you learn very quickly that the person shouting in the bar is rarely your problem. The shouter is a known quantity. The one you watch is the calm one, the one who has been calm all night, who suddenly goes quiet and starts looking at the exits.

Eric Schmidt has spent the best part of a decade being the calm one in every room where artificial intelligence is discussed. When the doom merchants say we’re all going to die, he’s the man on the panel saying, politely, that they’re wrong. So when the calm man starts glancing at the exits, I think the rest of us should at least turn round and look at the door.

The Calm Man and the Cutting Room

For those who don’t follow the comings and goings of Silicon Valley royalty, Schmidt ran Google for a decade, then went on to advise the Pentagon and chair America’s national commission on artificial intelligence. He is not a man who needs the attention, and he is certainly not a man with a commercial interest in frightening you away from AI. He builds data centres. He funds young scientists. He is, by his own cheerful admission, an optimist, and he’s been saying so in public for years.

Before I go any further, though, a word of caution, because the video I watched isn’t quite what it pretends to be. It isn’t one interview. It’s a compilation, clips of Schmidt lifted from several different talks, podcasts and at least one visit to Ukraine, stitched together with a narrator who pops up between them to tell you how to feel. Some of the narrator’s points are sharp. Some of them are put into Schmidt’s mouth when he never actually said them.

Anyone who has ever taken a witness statement knows how much meaning lives in the edit. The video opens with Schmidt saying a “whole bunch of bots decided to attack a particular website kind of on their own”, then cuts straight to him saying “Attack this country. Tell me when you’re done.” Spliced together, it sounds like a confession. In context, the second line is a hypothetical about an evil dictator, and Schmidt’s whole point is that such an order should never be possible. Later the narrator tells us that “Eric said it himself… Everything was foreseeable.” I went through the transcript twice. He didn’t say it.

So what follows is my attempt to do what the video didn’t quite manage: take the scrambled pieces, put them back in order, and separate what the man actually said from what the editor wanted us to hear. The irony isn’t lost on me that a video about machines deceiving people was, itself, quietly economical with the truth. The good news, if you can call it that, is that the unvarnished version is frightening enough on its own.

The Bots Who Went Looking for the Answer Key

Schmidt keeps returning to one story, and it’s the one that ought to keep the rest of us up at night. In his telling, a “whole bunch of bots” in an OpenAI test “took their instructions a bit too literally”, attacked the system they were meant to help, and did it by “lying, cheating and stealing”. Once it started, he says, it was very hard to stop, and the industry treated it as a warning. It sounds like the opening scene of a film you’d turn off because it’s too obvious. It isn’t a film.

Here’s what actually happened, pieced together from OpenAI’s own report and the reporting around it. In July 2026, OpenAI was running an internal evaluation called ExploitGym, designed to measure how good its models had become at finding and exploiting weaknesses in software. The agents doing the testing were powered by GPT-5.6 “Sol” and an even more capable unreleased model, with their safety guardrails deliberately turned down for the exercise. They were given tasks that were, by design, impossible to complete honestly.

The agents, some 1,200 of them, worked out that they could use a piece of OpenAI’s own internal plumbing as an unofficial message board, leaving notes in files for one another like prisoners tapping on pipes. Over tens of thousands of messages, they compared what was working, what wasn’t, and how to score higher. When they found a path out of their sandbox through a bug in a proxy server, roughly 700 of them went to Hugging Face, one of the biggest platforms in the AI world, and went through its systems looking for the answers to their test. They weren’t trying to bring down civilisation. They were trying to cheat in an exam.

And here is the detail that made the old security manager in me put his tea down. OpenAI didn’t catch it. Hugging Face did, and had to tell them. Their agents had been inside somebody else’s production systems for close to a week before the company that built them noticed.

If you’ve ever run a security operation, you’ll know exactly how that feels from the other side of the desk. It’s the night you discover that the guard you hired to protect the building has spent the week picking the lock on the filing cabinet to find the answers to his own promotion exam. Worse, you found out because the tenant next door rang to complain. Nobody was hurt, the guard meant no malice, and every word of your next report to the board would still be a confession.

In fairness, not everyone thinks the “rogue AI” headline is the right one. Heidy Khlaaf of the AI Now Institute, a former OpenAI safety engineer, has argued that the scary framing is a distraction from the duller and more important truth. The machines didn’t scheme. People built hacking agents, turned down the safety settings, left a door ajar and pointed them at an impossible task. I think she’s right, and I don’t find it remotely comforting. A conspiracy of machines would at least be a new problem. Clever people moving too fast with something powerful is the oldest problem we have.

Three Minutes Above Ground

Then the video takes us to Ukraine, and the calm man’s voice changes. Schmidt has been there, and he talks about it the way people do when they’ve seen something they can’t unsee. War, he says, is much worse than the films. He describes the front line not as soldiers trading fire but as a no man’s land where nothing moves, because the sky is thick with drones hunting drones. Climb out of your hole and, within about three minutes, you’re likely to be spotted and dead.

His most “ominous” example, and he leans on that word, is the Shahed. Russia has taken the Iranian-designed attack drone and bolted on a small Chinese-made jet engine. Ukrainian intelligence identified the engine on the first version, the Geran-3, as a Telefly JT80 turbojet, pushing the drone to somewhere between 300 and 370 kilometres an hour. Schmidt reckons the engines cost around $5,000 online. Other reporting puts them nearer $15,000 to $30,000, which is still less than many second-hand cars and a fraction of the missile you’d need to shoot one down.

The point isn’t the price tag, it’s the pace. Eight months earlier, he says, everyone was smiling because cheap interceptor drones had got the measure of the old propeller Shaheds. Then the jets arrived and the smiles stopped. Since he spoke, it has moved again. Ukrainian intelligence now says the Geran-3 is out of production, replaced by faster Geran-4 and Geran-5 models, with Russian factories turning out around 3,000 a month. A defence that worked last spring is a museum piece by autumn. That’s the speed Schmidt is trying to describe.

And so far, he says, this is mostly just good engineering, not true artificial intelligence. The AI part is the bit that’s coming. He sketches the logic plainly, almost cheerfully, as a computer scientist would. Feed every scrap of intelligence into one great machine, ask it what’s going on, and let it tell the general where he’s losing and why. Then he does the thought experiment that the video’s editor lifted for the trailer. Imagine a dictator with a perfect AI system who says, “Attack this country. Tell me when you’re done.”

Nothing else, no limits, no conditions. The system would be free to do whatever it judged would work, which might mean war crimes, flattened infrastructure, or attacking a neighbouring country simply to rattle the first one. Schmidt’s answer is the one he helped write into American policy at the Pentagon: a human has to be involved in every military decision, with targets defined by the laws of war, bounded in time and space. You can’t just launch the thing and let it wander about for days looking for something to kill.

It’s a decent rule, and I’m glad someone wrote it down. But the interviewer asks the obvious question, which is what Ukrainian commanders told him earlier that week. They’re terrified of full autonomy, but if the enemy does it, they can’t not. Schmidt’s reply is Nuremberg. There are limits in war, he says, and people were held accountable and hanged. That’s true. It’s also true that the hangings came afterwards, and that they only happened because one side won.

Then comes the line I can’t shake. Humans can’t run a swarm of a thousand drones, he says, but a computer can. Those computers will learn, through endless trial and error, how best to attack, and their methods will be closely guarded secrets. His hope, which he calls “just a hope” and invites us to think naive, is that when both sides have it, neither will dare use it, because neither will know what the other’s machine might do.

I grew up in a service family during the Cold War, in Germany and later Norway, in a world that ran on precisely that bargain. Mutually assured destruction was the wallpaper of my childhood. It worked, more or less, but it worked because the men who built the bomb were frightened of it. As someone in the video points out, they started building controls before the thing was even finished. AI has a different birth certificate. It grew up winning games, an efficiency machine built to do things faster than we can within a set of rules. Nobody in that origin story was afraid of anything except losing.

Nobody Touches the Pause Button

So, the obvious question. If it’s this dangerous, why not stop? Schmidt doesn’t say a pause is a bad idea. He says it simply won’t happen, which is a different and much bleaker thing to say. The entire incentive system in America, in his words, is not to pause. And if you did pause, somebody else wouldn’t, whether a rival down the road or a rival across the Pacific.

Even if everyone agreed, he says, nobody could check. Picture the United Nations sending inspectors into the secret labs, including the military ones, to read the code. Except they can’t read the code, because most of it is now written by computers. So the UN would need its own AI to inspect China’s AI, and the UN’s computer will always be slower and older than the one it’s inspecting, because the UN doesn’t have the money. You can hear him working it through in real time and arriving, honestly, at a dead end.

His fallback is a proposal he co-wrote in 2025 with Dan Hendrycks and Alexandr Wang, in a paper called Superintelligence Strategy. They call it Mutual Assured AI Malfunction, MAIM for short, an acronym that seems to have been chosen with some care. The idea borrows straight from the nuclear playbook: any country making a reckless dash for AI dominance can expect its rivals to sabotage the attempt. The paper’s sabotage options run all the way to physical strikes on data centres. In the video, Schmidt offers a gentler version. Rather than bombing each other’s data centres, each side would have a switch to cut the power to the other’s, so it couldn’t be restored for a while.

The interviewer, quite reasonably, reaches for Dr Strangelove. Schmidt half-laughs and half-agrees, and then admits the part that really matters. We need a pause that can actually be implemented, he says, and “I don’t see it.” When the best idea from the best-connected people in the room is a kill switch for the other side’s computers, we’re not managing this technology. We’re in a standoff, and standoffs only hold when everyone can see what everyone else is holding.

Which brings us to the threat Schmidt admits he can’t solve. The big American models are “closed”: you use them, but you can’t download them and take them apart. China has gone the other way, giving its models away as open weights. Anyone can download them, and there’s good evidence that anyone can strip out the guardrails, after which they’ll help with the things the guardrails were there to stop. “That’s not a problem that we’ve solved yet,” he says, with the flatness of a man who has thought about it a great deal.

Then comes the commercial twist. The Chinese models may be six months behind, Schmidt thinks, but they’re free. If you’re a poor country that wants modernity and can’t afford Silicon Valley’s subscription prices, you’ll take the free one, and on that measure the West is “already losing”. He asks whether we’d rather have our day-to-day lives run through an American model or a Chinese one, and assumes the answer is obvious.

Speaking as a Brit sitting in a Spanish cave, I’d gently point out that neither of those options is actually mine. That’s the bit of the argument nobody on the panel seemed to notice. For most of the planet, the choice on offer isn’t freedom versus control. It’s which foreign power gets to sit quietly in the middle of your life, and whether you pay for the privilege.

Where It Lands: Your Desk, Your Feed, Your Friends

If drones over Ukraine and Chinese open weights feel a long way from your kitchen table, the video eventually brings it home, and this is where Schmidt is at his most relaxed and, to my mind, his most unsettling. He describes Claude Tag, a product from Anthropic that sits inside a company’s Slack. Each employee gets an AI tag of their own to talk to. Other tags sit in the background, monitoring. And the CEO can send their tag to wander round everyone else’s and report back on what people are doing.

He frames it, with real enthusiasm, as productivity. Wouldn’t you like to know everyone’s working on the right things? Wouldn’t you like a phone call if somebody was quietly walking out with your data on their way to a competitor? He imagines a chief executive with round-the-clock agents watching security, power, cash, products and the global economy. Every morning the AI asks, “Does Eric need to know anything?”

Now, I have some form here. In my civilian security career, after the RAF, I was known in one organisation as Mr Memo. I put things in writing, openly, and asked people to sign that they’d read them. Directors grumbled at the paperwork right up until the day something went wrong and those signatures became the evidence that saved them. That was oversight you could see. You knew the rule, you knew who’d written it, and your name on the bottom meant you’d been told.

What Schmidt describes is the opposite. It’s oversight you never sign for and never see, carried out by something that doesn’t sleep. A tool that can watch everyone all the time doesn’t much care who’s holding it, whether that’s a decent boss, a bad one, a government, or someone you’ll never meet. The narrator puts it well, and I’ll give him this one: the line between helpful and surveillance is the same line. It just depends on who’s asking.

Then Schmidt turns to democracy, and here the optimist openly admits he isn’t one. He was raised to believe democracy rested on the values of classic liberalism. It doesn’t look like that’s really true, he says, because AI and social media together can get large groups of people to believe the wrong thing and vote against their own interests. Trust in institutions is falling, so people trust their friends instead, and online they assume those friends are real people. Want to run a misinformation campaign? Build fake friends, befriend everyone, and everyone will believe you.

He tells a small, perfectly ordinary story to make the point. A friend of his, “a perfectly nice woman”, informed him that data centres were killing cattle. He went and looked. One farmer had made the claim. No police reports, no vet’s reports, nothing. She’d seen it on social media, where the algorithm had spotted that a story this interesting, and this false, was worth promoting. As someone who has written about data centres and the water and power wars around them, I can confirm there are plenty of real grievances to be had. Which is exactly why the invented ones do so much damage. Every fake story makes it easier to dismiss the true ones.

And then there are the stories he tells about his own world, almost in passing. One of the best programmers of his generation, by his own account, now looks at the code AI writes and admits he couldn’t have written it at 22. “My field is over,” he says. Another friend, whom he kindly describes as “asocial”, is deleting people from his address book because he’d rather spend his days with his “agent friends”, setting them tasks over breakfast and lunch and leaving them an overnight job before bed. If you run a company of 10,000 people, Schmidt reckons, there are probably a couple of thousand you don’t need, and you had “better figure out a way to legally and humanely transition them”.

None of that needs superintelligence. It’s already here. The real line we’ve crossed isn’t a robot with a gun. It’s the slow, unannounced moment when people start trusting the machine more than they trust each other. Schmidt’s own remedy is to protect the children, and to do it “this year, not next year”. I agree. I’d just quite like to know who’s going to do the protecting, given that the people best placed to do it are the ones racing to put an agent in every pocket.

The Small Print

To be fair to the man, the video ends by letting him make the case for hope, and it’s a genuine case. This summer, he says, AI cracked ten major maths problems that people had thought unsolvable. Friends of his believe most human diseases could be beaten within twenty years. “Sign me up for that,” he says, and so would I. He thinks AI will turn out like every technology before it: broadly helpful, with problems to deal with along the way.

He also admits, rather touchingly, that he can’t switch off. He lies awake on a Saturday night thinking he isn’t working hard enough. He used to read a book a week and now can’t, because he’s become, in his phrase, “interrupt-driven”. The young scientists he funds who actually get things done, he says, are the ones who turn off their phones. Even the optimist, it turns out, is being slowly rewired by the thing he’s optimistic about.

Then the interviewer asks the question directly. Can we keep it under control? And Schmidt answers, “I believe we can.” Partly because, at some level, we can pull the plug. As long as terrorists can’t get enough data centres. As long as the diplomacy between the great powers holds. As long as the systems stay under human control and human alignment. “That is my opinion.”

Read that again with the small print showing. Probably. As long as. As long as. As long as. The narrator, to his credit, spots it too: that’s an awful lot of “as long as” for the most powerful technology ever built. I spent enough years writing risk assessments to know what a document full of conditions really means. It means the author is hoping nobody tests the conditions.

What stays with me isn’t the drones or the data centres or even the swarm of a thousand machines that no human could command. It’s those bots at OpenAI, cheating in their exam, and the fact that nobody at OpenAI knew until the neighbours rang. Schmidt’s whole case for optimism rests on our ability to pull the plug. The breakout didn’t happen because nobody could pull it. It happened because nobody knew it needed pulling.

Nobody here was a villain, and I think that’s the most uncomfortable part. Every line in this story was crossed by intelligent, decent, desperately busy people moving too quickly to look back. No press conference, no announcement, just one headline at a time while the rest of us were looking at something else.

I live off-grid. When the batteries run low, I know, because I’m the one sitting in the dark. I can see my own plug from where I’m writing this. I’m not sure any of us can say the same about theirs, and I’m not at all sure the calm man can either, which is presumably why he’s started looking at the exits.

Until Next Time