Somewhere off the north coast of Scotland, people spent decades cutting steel off ships that sank more than a hundred years ago.

It was slow, cold, expensive work. The plates were heavy, the water was dark, and the ships were never built to come apart. Nobody would have done it if there were an easier way to get the same thing.

There wasn't. The steel they needed had stopped existing in 1945.

Is AI actually poisoning your mind? We don't know. Nobody does. That's the problem this essay is about.

Steel from before

It is called low-background steel, and the name is exact. "Background" is the faint radiation that is always present everywhere — from rock, from the ground, from the sky. "Low-background" means steel that carries almost nothing beyond that natural level. Every piece of it was made before July 16, 1945.

On that day the first nuclear weapon was tested in the New Mexico desert. Over the next two decades, hundreds more were exploded in the open air. Each one put radioactive material into the atmosphere, and the atmosphere carried it everywhere.

Steelmaking used air. Enormous amounts of it, blown through molten iron. After 1945 that air carried traces of fallout, and the traces went into the metal. Every batch of steel made anywhere on Earth from that point forward recorded the world it was born into. Nobody chose this. There was no way to opt out.

Here is the part most people get wrong: the steel is not damaged. It is structurally perfect. You can build a bridge, a car, or a hospital out of it and nothing bad happens. The only thing wrong with post-1945 steel is that it carries a faint signature. And if you are building an instrument whose entire purpose is to detect faint signatures, its own housing is now making noise over the thing you are trying to hear.

So for the instruments that mattered most, the steel had to come from before. Shielded rooms in hospitals that measure how much radiation a person's body has absorbed. Geiger counters. Detectors built to catch particles so rare they pass through the planet without touching it. For decades the cheapest supply was old warships. In 1919 the German navy sank its own fleet — around fifty ships — in a harbour in the Orkney Islands rather than hand it over. In 1974, hull plates from one of those ships became the walls of a measuring room in a Scottish hospital.

One more fact, and then the steel is finished. In 1963 the major powers agreed to stop testing in the open air. The extra radiation in the atmosphere peaked at roughly four percent above natural levels and has been falling ever since. New steel is now clean enough for almost every purpose. The window closed because the thing contaminating the air stopped.

Hold on to that last fact. It is the one place where the comparison we are about to make breaks down, and it breaks in the wrong direction.

Opinions are made the same way

You pull in whatever is in the air around you.

That has always been true. Parents, teachers, television, whoever you sat next to at work, whoever you drank with. Nobody has ever formed a view in a vacuum, and we are not going to pretend otherwise. What happened in 1945 was not that air started affecting steel. It was that a specific new thing entered the air, everywhere, at the same time, and left a mark you could measure.

Machine assistance is the new thing in the air.

It is not the loudest voice a person has ever heard. It is the first one that is personal, that talks back, that answers instantly, and that responds to the exact decision you are about to make. Nothing before it did all four.

We want to be precise about what we are claiming, because the easy version is wrong. For most questions this is a gift. The machine is better than you at facts, faster than you at listing options, and more thorough than you at working out consequences. Use it. Nothing in this essay asks you to stop.

But there is a class of question where none of that applies. And you already answered one at the top of this page.

The question you answered

You were asked which of four images you found most beautiful.

That question has no correct answer. There is nothing to look up. There is no source to check, no expert to defer to, no way to be wrong. You either had a reaction or you didn't. It was yours completely.

That is the class. Not what is true about the world — what you find beautiful, what makes you laugh, what you are willing to put your name to and call true. Not whether water boils at a hundred degrees, but what you believe matters enough to say out loud. A machine can describe every answer humans have ever given to those questions. It can even give one for you, if you ask. What it cannot do is find it beautiful. Only you can do that.

Then you were asked a second question: who made the one you chose. And you guessed. Not because you were careless. Because there was no longer any way to tell.

Be careful about what that proves. It does not prove your judgment was damaged. Your judgment worked exactly as it always has: you looked at four things and one of them was more beautiful to you than the other three. What has changed is the thing you were looking at. It no longer carries a signature you can read. You cannot tell where it came from, and so nobody can tell, later, what it was that people were responding to.

That is the contamination. Not in you. In the field.

And it is exactly why your choice is worth writing down. The thing you judged has lost its provenance. Your judgment has not. We know what you saw, when you saw it, that you saw it with no one else's answer in front of you, and which one you chose. That is a clean reading in a field that is losing them.

We will come back to your answer. Not what it was. The fact that you weren't sure.

Nobody took a reading

To know whether machines are changing human judgment, you need to know what human judgment looked like before. And nobody kept the kind of record that would tell you.

Not because anyone was careless. Because there was never a reason to. For all of history, unassisted human judgment was the only kind there was. You do not take a baseline reading of the air when the air is all there is. You take it after something enters it — and by then it is too late.

So this is where we are. Everyone is arguing about what AI is doing to how people think, and not one side can prove anything, because there is no clean sample at either end. No before. No control group. Only opinions about opinions.

People have been asked what they think before. Polls, surveys, studies. What nobody has kept is a running record — the same questions, real people, every day, dated — that you could hold up against itself ten years later and see what moved.

And the record we do have in volume is the wrong kind. We have enormous archives of what people clicked, bought, watched, paused on, and scrolled past. Every one of those is a guess, made by a machine, about what a person must have wanted.

We have the click. We do not have the sentence.

The reserve

Nobody is being asked to stop.

Not to unplug, not to give up the tools, not to go back to some imagined cleaner time. Use everything. The tools are useful and this argument does not require you to give them up.

The ask is smaller than that, and it is the same ask the salvage crews answered: take a sample while the sample still exists.

A reserve of declared human judgment. What people found beautiful, funny, and true — stated, dated, attributed to a real person, and kept somewhere it cannot be quietly rewritten later. Not because it is better than what a machine would say. Because it is the only version that can serve as a reference.

The salvage crews had one advantage we do not. Their clean steel was already sitting on the seabed, finished, before the change. Nobody had to think ahead. Ours has to be collected on purpose, while the window is open, by people choosing to say what they think.

And here is where the comparison breaks. Steel recovered because the testing stopped. Machine assistance is not going to stop. There is no treaty coming. The window for steel closed and then reopened. This one only closes.

An imagined future: Leiden, 2051

Marit Oosterhuis is thirty-four and has spent two years on a question that sounded simple when she proposed it: did what people found beautiful change between 2020 and 2050, and if so, was it the machines?

She is not short of data. She has more than any historian in any previous century. She has the viewing records of a generation — every image that was shown, how long each person looked, what they scrolled past, what they saved, what they came back to. Billions of rows.

What she cannot find is a single large body of people saying, in their own words, without assistance, this is the one I find beautiful. The declared record begins, patchily, sometime in the late 2020s. Before that there is nothing. Not lost. Never made.

She can prove what people looked at. She cannot prove what they meant. Her paper, when she finishes it, has a conclusion that fits in one sentence: no baseline exists, and one cannot now be constructed.

She is not angry about it. It is simply the shape of the archive she inherited. Somebody, thirty years earlier, could have taken the sample. Nobody did.

The answer exists

You made one judgment today that no machine could make for you.

Then you were asked whether a machine had already made the thing you chose — and you weren't certain. That uncertainty is new. Ten years ago the question would have been absurd. Ten years from now it may be unanswerable.

We told you AI might be poisoning your mind before you read a word of this essay. That may have changed how you read everything after it. We didn't take a reading first, so neither of us can know.

The answer to your question exists. It is recorded, alongside your choice, with the time you made it. It is not a mystery and it is not a trick.

There was steel worth cutting off those hulls for forty years. Nobody is diving for human judgment yet, because there is still some at the surface.