ai
A Watermark That Proves Almost Nothing
Since 2 August, Claude marks the text it generates. Anthropic is open about the limits: the mark cannot separate writing from editing, and its absence proves nothing at all. Which leaves exactly one use for it, and it is not a good one.
A watermark found does not mean the AI wrote it. A watermark missing does not mean you did.
Since the second of August, Claude marks the text it generates, and Anthropic says quite plainly what that mark is worth: it indicates that Claude was probably involved somewhere along the way, it cannot separate writing from heavy editing, and it says nothing about how much. When no mark turns up, that proves nothing either, since a different model leaves a different signature, or none at all.
So we now have a signal that cannot answer the one question anyone will ever ask it.
The reason is the EU AI Act. Anthropic is one of around 190 signatories doing what the Code of Practice asks, which puts the complaint in Brussels rather than San Francisco. Montesquieu set the test long ago: if it is not necessary to make a law, it is necessary not to make one. A rule that cannot deliver what it was written for fails that test twice, first on necessity, then on effect.
There is meanwhile a small market for removing a watermark that nobody outside Anthropic can measure yet, because the detector has not shipped.
What remains is the part I like least. The only practical use of a signal this weak is pointing at someone ... a pupil, a job applicant ... and saying: it was you. I struggle to think of anything less productive. Or, forgive the word, anything uglier.