A perfect AI detector wouldn't fix the internet

I was halfway down a recipe page, scrolling past four paragraphs about someone's grandmother's kitchen, when I caught myself wondering whether a machine had written it. The grandmother, the kitchen, the misty autumn mornings. It had that slightly too-smooth feel.

Then I realised the question didn't matter. The only thing I needed to know was whether the cake would work. And if it didn't, there was nobody on that page I could tell.

The worry, stated fairly

The dead internet theory, in its current form, goes something like this. The web is filling up with AI-generated content. Articles, reviews, comments, whole websites, produced faster than any person could write them. Eventually the real people get drowned out, and what's left is machines writing for machines, with humans wandering through a place that only looks inhabited.

It's not a silly fear. There really is a lot more synthetic text around than there used to be, and some of it is bad.

The usual responses come in two flavours. One says it's not that bad: most of what you read is still written by people, relax. The other says it's worse than you think: the machines are already the majority, and you can't tell. Those sound like opposite positions, but they agree on the important part. Both treat human versus machine as the line that matters. Both assume that if we could just tell which was which, we'd know what to trust.

That assumption is the mistake.

Label every page

Imagine that tomorrow, every page on the internet gets a small label at the top. Written by a person. Or Written by a machine. Perfectly accurate, impossible to fake, no exceptions.

That's perfect detection, the thing the whole worry says it wants. So, before getting too excited: what would actually get better?

The scam site written by a real person still takes your money. It just has a label on it saying a human did it. The fake review typed by someone paid to type fake reviews is still fake, and it's now officially certified as human. The misleading article written by a person who didn't believe a word of it is still misleading.

Meanwhile, the accurate machine-written summary of a train timetable is still accurate. The clear instructions for resetting a router are still clear. The recipe still works or it doesn't, and now there's a badge at the top telling you a machine was involved, which tells you precisely nothing about the cake.

Sort the whole web into two perfect piles and the useful stuff and the rubbish stay mixed through both of them. That's the test of whether a worry is framed properly. If solving it completely leaves the thing you cared about untouched, it wasn't the thing you cared about.

Machines didn't invent hollow text. They made it cheap.

The web was full of text nobody meant long before any machine wrote a sentence of it.

For years, a lot of what turned up in search results was written by people paid by the word to fill a page around a keyword. Reviews were written in bulk. "Personal" posts were written by someone whose job was to sound personal. None of it was machine-made. Plenty of it was every bit as hollow as anything a language model produces now.

Nobody called that a dead internet, because a person had typed it. Which suggests the human origin was never what made text worth reading. What machines changed is the price. Hollow text used to cost something to produce, and now it costs almost nothing, so there's a lot more of it. That volume is a real problem. But it's a problem of quantity, not of category, and the fix for too much of something has never been checking what species made it.

The question that actually matters

So what were we worried about?

Go back to the recipe. The trouble wasn't that a machine might have written it. The trouble was that if it was wrong, nobody would ever know, fix it, or be embarrassed about it. There was no name, no editor, no one who'd put their reputation on the cake rising.

That's the real variable, and it's been sitting under the dead internet worry the whole time. Who stands behind it, not who wrote it. A machine-written summary that a named person has checked, and will correct if it's wrong, deserves more trust than a human-written column that nobody stands behind. A page with an owner can be wrong and get better. A page without one can only be wrong.

When a piece of writing makes you uneasy, the unease usually isn't "a machine did this". It's "if this is wrong, nobody's coming". The costume changed. The worry underneath is old.

A return address

We never really trusted words because a person wrote them. Newspapers ran unsigned pieces for generations. Speeches were written by someone other than the speaker. Company reports were written by committees nobody could name. People trusted them anyway, or didn't, based on something else.

The byline was never proof of a person. It was a return address. It told you where the blame would land if the words turned out to be wrong, and that someone there was prepared to receive it.

The internet isn't dying because machines have started writing. It's filling up with letters that have no return address. Some of them were written by machines. A lot of them weren't. And the label that would tell us which is which was never going to be the one we needed.