AI Watermark

It's embedded at the moment of creation

A watermark isn't painted on after the fact — it's worked into the generation process itself. An image model folds in a tiny, rule-based bias at the stage where it's placing pixels; a text model does something similar, slightly favoring one word over other equally fitting choices, following a set rule.

Looked at in any one spot, the bias is indistinguishable from chance. Only when you gather it up across the whole piece of content and count it together does "a pattern too consistent to be chance" show up. That's also why detection can still work even after part of the content has been cut away.

Because it's added during generation, it can't be added afterward by someone else. Trying to plant an AI watermark onto someone else's existing photo, to falsely claim "this was AI-made," is difficult without knowing the embedding rule.

It only works if it stays invisible

A watermark has to satisfy two goals that pull against each other: the less noticeable, the better, and the more it survives rough handling, the better. But these move in opposite directions. Embed it strongly and it lasts longer but degrades quality; embed it lightly and it stays invisible but washes out easily.

So it gets placed where people are least likely to notice — fine texture in color rather than plain brightness, busy stretches of sound rather than quiet ones. It hides the mark in the gap where human senses are least sharp.

Only someone who knows the rule can check

Nobody can read an embedded pattern just by looking. You need a detection tool built by whoever knows the rule and where it was scattered, in order to say "this piece of content carries a mark." Even then, the result isn't a plain yes or no — it comes back as a confidence score.

That score needs a cutoff line, and drawing one brings two kinds of mistakes along with it. Set the bar low, and human-made content gets mislabeled as AI-made; set it high, and AI-made content slips through undetected. Where the line gets drawn depends on which mistake is more costly.

If every maker uses a different rule, whoever is checking ends up needing to carry around several different tools. There's ongoing discussion about aligning these methods, but nothing has converged into one standard yet.

It can be erased

This is the most important limitation. Cropping, resizing, re-saving at lower quality, filming a screen that's displaying it — the mark chips away a little with each of these. Tools built specifically to target and strip the mark exist as well.

No mark doesn't mean a human made it. It might have been made with a tool that doesn't embed one, or it might have carried one that faded during copying. A watermark is fairly reliable in the direction of "if it's present, it's AI-made" — but nearly useless in the direction of "if it's absent, it's human-made."

It's harder to embed in text

Images and audio have plenty of room a person won't notice. Text has far less of that room. Change one word and the meaning shifts, and a short sentence barely has enough space to scatter a pattern across in the first place.

Text is also far easier to edit. Touching up a few sentences, or running it through another tool once, is often enough to scatter an embedded bias. That's why detection holds up reasonably well in long text but is hard to count on in short text.