Look closely. Gaze upon the yellow protrusion and the dark bags beneath the eyes in the portrait of Rembrandt in Kassel. Or perhaps Fra Filippo Lippi’s tightly verticillated ears, or Giovanni Bellini’s prodigiously long thumbs. This was the occupation of the self-appointed art critic Giovanni Morelli, who, dismayed by the state of nineteenth-century art criticism, devised a new way to investigate a painting’s attribution.

For Morelli, the true signature of an artist’s hand resided in the small details. A skilled copier might naturally devote his attention to what viewers would notice first. Morelli looked instead at what the artist rendered almost without thinking: “earlobes, fingernails, shapes of fingers and toes.” Such details were produced, as Carlo Ginzburg put it, “by force of habit, almost unconsciously.” They were the characteristic trifles by which an artist gave himself away.

When Anthropic announced in August that future Claude models would watermark their text, it revived a similar hope: that a piece of writing might carry some minute evidence of the author behind it. It also revived a familiar anxiety. Would the mark make Claude’s already recognisable prose more stilted? And what would it mean to mark a passage as Claude’s if–indeed it was a person– who had prompted it, revised it, or supplied most of its words? In a classroom or a court, the difference between wrote, edited, and was involved can matter enormously.

Anthropic’s own example to demonstrate its method uses the sentence: “The weather today was cold and…”. The next word is unlikely to be “sugary,” but it could plausibly be “overcast” or “grey.” Between these two words, Anthropic contends, very little is at stake and so the model would ordinarily settle the choice between these two words with a random number generator. With watermarking, a secret key combined with the preceding words is used to settle the choice over randomness, and is thus completely coded and checkable against this key.

A recent Nature paper describing the method Anthropic uses makes a revealing qualification: “Tournament sampling performs better when there is more entropy in the LLM distribution, and is less effective when there is less entropy.” Entropy, here, means how much choice the model has at a given point. If several next words are plausible, entropy is high. If one word is overwhelmingly likely, it is low. “Performs better” means that the watermark is easier to detect.Watermarking therefore measures unrestricted choice, and reports authorship as a by-product. Arbitrary wording then becomes the measure of detectability.

This may seem contradictory. Training makes machine text more recognisable as machine text, hence the emergence of a ‘house style’, as it were. But a house style identifies a type. It cannot establish that a particular passage came from a particular model. The watermark attempts that narrower task. The same pressure that produces a recognisable style may leave less room to mark an individual passage.

So this begs the question: are heavily trained registers – refusals, safety language, the standard forms of response to someone in distress – harder to watermark? If those responses really do converge on a smaller range of wording, the passages that sound most carefully composed to express care might carry the least detectable signature. I have yet to see detection rates measured by response type.

And there is another difficulty. Anthropic says light editing of generated text may leave the watermark intact, while a complete rewrite can remove it. Conversely, if Claude merely proofreads a human passage, there may be too few choices of its own for the mark to register. The visible trace and the degree of influence need not move together. A machine could plausibly alter the course of an argument with a suggestion the writer later puts wholly into her own words.

Will the watermarks dissolve? Some will, like ink siphoning itself into a current. But perhaps the more interesting question is what we expect to find if they remain. The Kassel portrait that invites us to look so closely is catalogued today as a work from Rembrandt’s workshop: his face, though apparently another painter’s hand. With machine text, we may have to look in an unexpected place. If the most recognisable habits carry the weakest signature, what, exactly, are we learning to recognise?