No, AI is NOT going to kill us.

This week it was Anthropic’s turn. One of its researchers quit his job and then said publicly what he had apparently been worrying about privately for some time… that the AI his former employer is helping to build could wipe out humanity, possibly before 2030.

His post did more than a hundred million views, which is pretty good going for something that can be loosely summarised as “we might all die, and I was involved”.

Then it got stranger. Some of his former colleagues more or less backed him up. One said, with very little room for misinterpretation, that they genuinely believed AI could kill everyone. Axios has the full story. Bernie Sanders is now talking about banning superintelligence altogether. So yes, you may have heard something about it.

I wrote a novel about roughly this earlier in the year. Ghosts and Gods is, among other things, about artificial intelligence going badly wrong. While I was writing it, I spent months reading AI safety papers, government reports, extinction arguments, counterarguments and the increasingly popular genre of resignation letters written by someone leaving an AI company because they believe they may have seen the end of civilisation coming. I wanted the book to feel plausible. Not necessarily likely, but plausible enough that a reader might occasionally stop and wonder whether we were stupid enough to end up there.

The funny thing is that the more I researched, the less frightened I became.

Let’s try something quickly. 👇 Look at this image. Do you see a robot dog, or a killing machine? If it’s the latter, that’s your bias at play, not the machine’s.

So, are the machines going to kill us?

No. They won’t. I realise this is badly timed, given that people in Silicon Valley are currently attaching percentages to the likelihood of our extinction as casually as the rest of us discuss the weather. But once you get beyond the people giving their personal “p(doom)” and into the boring business of how an AI might actually kill everyone, things get considerably less cinematic.

RAND looked at this properly. They examined three routes an advanced AI could theoretically use to cause human extinction: nuclear weapons, engineered pathogens and climate manipulation. Then they worked through what would actually have to happen for each scenario to succeed. It is grim reading in places, but also strangely reassuring. Their conclusion was not that extinction was impossible. They were careful not to say that. Their conclusion was that actually pulling it off would be extraordinarily difficult. The full RAND study is here.

That distinction matters. “We cannot completely rule this out” is not remotely the same statement as “there is a good chance most of humanity will be dead within four years”. Yet somewhere between those two sentences, an enormous amount of the public AI debate seems to be taking place. One is the language researchers use when they do not want to pretend certainty where none exists. The other gets you invited onto podcasts.

The bit I struggle with

The extinction argument usually requires us to imagine a particular sort of AI. Not just something vastly more capable than what we have now, but something with its own objectives. Something that decides human beings are interfering with those objectives and begins acting against us.

That is the point where I start to lose the thread, because if you go looking for evidence that today’s systems are developing anything resembling independent wants, motives or ambitions, there really isn’t much there. Arvind Narayanan and Sayash Kapoor at Princeton put this better than most in their essay “AI as Normal Technology”. Their argument is not that AI is unimportant. Quite the opposite. They expect it to be enormously consequential. But they suggest thinking about it more like electricity or the motor car than the arrival of a new species.

Transformative technologies spread unevenly. People use them badly. Companies oversell them. Governments misunderstand them. They create new industries, destroy others and produce consequences nobody planned for. That is quite enough to be getting on with without assuming the technology is also secretly forming an opinion of us.

Narayanan and Kapoor are not dismissive about AI risk. They just refuse to give a sophisticated tool a personality before there is evidence it has one. I find that persuasive.

The people worry me more

There are plenty of reasons to be worried about AI. I just think most of them still have a person attached. The gun lobby has been saying some version of this for decades, and while I am reluctant to give them the satisfaction, the basic grammar is useful: tools do not need motives if the people using them already have one.

The International AI Safety Report, chaired by Yoshua Bengio and produced by around a hundred experts with backing from more than thirty governments, divides the major risks into three broad categories. One of the biggest is malicious use. Fraud. Cyberattacks. Deepfakes. Manipulation. People producing lies and impersonations at a scale that would previously have required money, skill, staff or all three. The 2025 International AI Safety Report is here.

None of this belongs to some distant future. It is happening now. Stanford recorded 362 documented AI incidents in 2025, up from 233 the previous year. The figures are in its 2026 AI Index.

What interests me is how few of these stories involve an AI independently deciding to hurt somebody. Usually somebody wanted money, influence, access, revenge or attention. AI simply made the job easier. I studied cyberpsychology some years ago, and I think that probably colours how I look at all of this. When technology causes harm, my instinct is to look for the person first.

Somebody wrote the phishing message. Somebody cloned the voice. Somebody decided that pretending to be an elderly woman’s grandson was a reasonable way to spend a Tuesday afternoon. AI gives that person reach. It gives them speed. It can make a fairly mediocre criminal considerably more effective. But the intent was already there.

This is not particularly comforting

I’m not saying everyone should relax. Quite the opposite. I think we may be spending so much time worrying about the most spectacular possible version of AI risk that we are neglecting the boring version already causing trouble.

If every camera is pointed upwards waiting for superintelligence to appear over the horizon, nobody is watching the bloke three streets away running voice-clone scams from his spare bedroom. And there is another, less exciting problem. A lot of the behaviour we are frightened of is already illegal.

We do not need to invent an entirely new philosophical category of crime because somebody used AI to impersonate your daughter and steal five grand from you. The fraud is still fraud. The problem is catching the person doing it. That seems to me a much more immediate problem than drafting laws for a hypothetical machine that might one day decide it dislikes us.

There is also something slightly odd about the ritual of leaving an AI company and then warning the rest of us about what the people still inside it are building. I understand why someone would leave on principle. There are circumstances where that is exactly the right thing to do. But if you genuinely believe a technology has a meaningful chance of killing everybody, I am not sure walking out the door necessarily improves the odds.

The people with the best chance of reducing whatever risks exist are, for better or worse, the people actually building the systems. I would rather some of the worried ones stayed in the room.

Which brings me back to the crooks

Ghosts and Gods is still a novel about AI going wrong, and I would not change that. But if you read it closely, the machine is not really the thing to be frightened of. The people are.

That was not entirely where I expected to end up when I started writing it. I went into the research looking for material that would make the book scarier. Instead, I kept finding human beings. Human ambition. Human greed. Human stupidity. Human beings convincing themselves that whatever they were doing was justified because somebody else would do it if they didn’t. All the usual stuff, in other words.

So the next time an AI researcher resigns and announces a new percentage chance that humanity is finished, I will read it. Of course I will. I am fascinated by this stuff. But I will probably spend more time worrying about the person cloning voices, the person automating fraud, and the person creating a convincing video of a politician saying something they never said.

Because those people are not theoretical, they are already here. And while the rest of us are staring into the sky waiting for the machines to arrive, they are getting on with it.

On a slightly more practical level, if you want your AI-infused digital strategy to still be alive by 2030, do get in touch.