I repeatedly heard the term watermarking applied to AI output text and I was struggling to wrap my head around it. 

I mean, I get it. I get the purpose. Like when we watermark a photograph with a logo in the corner or a semi-transparent “PROOF” splattered across the middle to prevent someone from making a nice print from it, it all is there to denote both ownership and origin. The watermark is a kind of legal lock on a piece of content in that way.

But when talking about generative content we’re both (a) leaning more so into tracking the origin of the content (as not to confuse it with human-written words) and (b) no longer talking about something so obvious as a stamp across a picture.

I dug into this a little bit and it is actually pretty clever… as least as much as I can infer.

We humans, presumably and if you were to run even this very text through such an analysis, write in in predictably random ways. English grammar has a solidly predictable statistical spread of, say as an example, the letter E versus the letter J. If you’re just writing, you don’t even think about it. It’s mostly just a quirk of basic English.

AI isn’t bound by such organic randomness.

I don’t know precisely how each AI is warping this math—that’s kind of the point, I think, no one does—but when text is generated by an algorithm these kinds of statistical truths of human writing are nudged off kilter. That essay, email or short story that AI barfed out for you will not only lack the beautifully rough edges of human creation, but it will be statistically biased to use (say, for example) a disproportionate number of Js versus Es. A very specific ratio, in fact, and probably not just for that combination of letters or phonemes or punctuation marks.

The result is that a simple calculation could be done (so long as the calculator knows what it is looking for) and when it sees those specific ratios extracted from a piece of text it can say either (a) this is random enough to assume a meat brain made it or (b) I recognize the proportion of Js to Ks and Rs to Zs and so on and yeah, that was written by Claude. 

We can’t see the watermark, but math can. And so long as it’s there we can check to see where the text originated.