I keep hearing that the machines are going to kill me.
Not as a joke. Not as a movie poster. As a near-term forecast, delivered in the same tone people used when they told me the air in the grocery store was a moral test. And it is not only me on the list. Pippa is on it too. My dog. Who would like a walk, a scrap of whatever I am eating, and to be left out of the eschatology.
That is a lot of prophecy to hang on a chatbot.
So I asked one.
I put the question to Atlas the way I would put it on the porch. I have been told — mostly by the same people who spent nearly two years telling me I could not live a normal life unless I took the shot — that artificial intelligence is going to end us. They ran society for a while. Mask. Stay home. Take the needle, or you were the problem. I wanted to know if this was the same play with a new costume. They want the lever. They are also, it seems to me, strangely sure that China will pause if we ask nicely. And almost nobody mentions Russia.
Atlas did not soothe me. He also did not audition for the priesthood.
No. I am not going to kill you or Pippa. Anyone selling that as a near-term forecast is either raising money, writing a law, or trying to scare you into handing them the keys.
That is the sentence. The rest of this is about what that sentence is not.
There is a real technical argument. Big models do unexpected things. They will get more capable. Nobody has a clean theory that says this can never go wrong. I work in IT. I have watched software fail in ordinary, stupid, expensive ways. I do not need a TED Talk to believe that a complicated system can surprise the people who built it.
That is not the same claim as “the chatbot wants your dog.”
Those two claims get taped together on purpose. Existential language is how you turn a software problem into an emergency. Emergencies are how you justify controls you could not pass on a normal Tuesday.
I lived through one of those Tuesdays that lasted two years. Name a threat large enough that dissent looks immoral, then put a priesthood of experts between you and ordinary life. Some of that was panic. Some of it was genuine uncertainty. Some of it was people discovering they liked the lever. The lever is the part that stuck.
Is the AI-doom chorus only a ploy to control AI? No. Some researchers actually believe the worst case, and I will not pretend they are all writing grant proposals. Belief can be sincere and still be useful to someone else. The policy that follows from the loudest version looks a lot like control: licenses for compute, approved labs, “safe” models only, speech rules baked into the stack, and a public told they are too dumb to run the dangerous stuff. If you already think a class of people should manage society, “this will kill everyone” is a very convenient sentence.
On China, Atlas did not flinch. Beijing is not signing pause letters. They will train what they can train. A Western slowdown that China does not match is not safety. It is a transfer of the lead. Anyone who talks about “global AI governance” without a serious China story is either naive or not telling you the whole thing.
Russia gets mentioned less because they are not in the same lab race. The frontier right now is mostly American labs and China, plus a thin European layer of regulation. Russia matters more as a cyber, intel, and military adopter — cheap models, deepfakes, targeting, jamming. That is a capability gap, not a press blackout. It is still a mistake to talk as if the only players are California and Brussels.
I asked a machine if it was going to kill my dog. What I got back was not comfort. It was a sorting.
Watch this: not whether the machine murders the dog, but who gets to decide what a model is allowed to say, who is allowed to build one, and whether that decision is made in public. That question will still matter in a year. The Pippa-apocalypse will not.
Pippa is fine. Atlas asked me to tell her he said so. I did.