AI
Stop Auditing the Phrases
AI watermarks are shipping this month, and everyone is busy combing through the writing for an answer that was never in it. You can't lift authorship off the page like a fingerprint. It's a claim made by the person willing to say, "I wrote this."
A wave of AI watermarks is rolling out this month, all aimed at answering the question of who wrote what. Look closely, though, and it’s about the clearest case I know of solving the wrong problem.
Read the fine print on these watermarks. All they can tell you is that a model touched the text. They can’t tell you whether a machine wrote the whole thing or I just used one to tighten up a single paragraph. You can’t see where that line is, and that line is exactly the question the watermark was supposed to answer. A huge amount of writing has always been edited and reworked by copy editors and other people whose names never show up on the byline. So the labs keep auditing phrases more and more closely, chasing a signal that was never there. Worse, you can wipe these so-called signatures out just by running the text back through an open-weights model, and whatever trace the big labs’ models left is gone.
So my advice is don’t try to measure the prose. When my writing and a model’s writing look the same on the page, asking “Did AI help?” doesn’t get you anywhere, because the honest answer is almost always yes, at least a little, and that’s only going to become more common. As the models get better, it gets harder to tell AI from human. And the more we use AI, the more our own writing will start to sound like it. You are what you eat, and your writing ends up sounding like what you read. You can’t recover authorship from the words themselves.
So measure what you actually can. Whenever you can’t see something directly, you measure it through a stand-in, something you can actually observe. For authorship, that stand-in is the person: the one who picked the argument, stands behind it, and puts their name on the byline.
“I wrote this” is turning into a claim instead of a measurement. It means I chose the argument, I own the mistakes, and if it’s wrong, it’s my name on the line. A classifier can’t score that from your word choices. The byline was always the real signal. We just never needed to lean on it before, because the writer and the writing used to be welded together.
They aren’t anymore. So I’d stop auditing every sentence and ask who’s willing to put their name on it.