12 September 2026 · 2 min read
What is data watermarking?
It marks a dataset rather than a piece of content, and it answers one question: whether your data ended up inside someone else's model or someone else's copy of your table.
Data watermarking marks a dataset — a table, a corpus, a pile of training text — rather than any single image or paragraph. It exists to answer one question: did my data end up inside your model, or your copy of my database? It says nothing about whether a particular sentence was written by a machine.
Marking a database
The older version dates to a 2002 paper by Rakesh Agrawal and Jerry Kiernan at IBM, and it is still the basic recipe. A secret key picks a scattering of rows, a numeric column in each, and a bit near the bottom of the number stored there. Those bits get set to values the key can regenerate later. Each change is small enough to sit inside the error the data already tolerates, so the table stays usable, but the pattern is far too specific to turn up by chance. If a licensed copy appears somewhere it should not be, the key finds the mark.
Marking a training set
The newer version aims at models. One approach is the copyright trap: plant unique synthetic sentences in your text, then later prompt a model and see whether it has memorised them. Another is to watermark the text itself, so the bias in its word choices survives training and shows up in what the trained model writes. Researchers call that radioactivity.
Both are research rather than products. Traps work by repeating a sequence many times, which is the exact pattern de-duplication strips out of a corpus before training starts — fuzzy variants that alter a few tokens per copy exist to survive that. Either way, running the check needs the model, or at least a lot of queries to it.
What a match would and would not show
A hit says your data was seen. It does not say who wrote that data, and it does not settle whether the use was allowed — that argument is a legal one either way.
It also does not work backwards. No mark in a dataset can tell you anything about the paragraph in front of you. For that, check the text for hidden characters, or read the explainer on AI watermarks for what those marks do and do not prove.
Check something yourself
Every tool here runs in your browser and is free. See what your files and text actually carry.