How far does the mark travel?
Anthropic said on Monday that its new Claude models embed a watermark readers cannot see into the text they generate. As Business Insider reports it, the mark leaves the meaning and readability of the text unchanged, travels with it when it is copied and pasted, and may persist through some editing. Models launched on or after August 2 carry the support from launch, and work continues on older models. The marking also applies when Claude is reached through cloud providers.[1]
The same account lists the limits. Heavy editing, paraphrasing, translating or mixing Claude's output with other writing can make the watermark undetectable. And finding a watermark does not show that Claude wrote the passage, because a model used only to proofread or translate can leave a mark too. The company says it plans to give third parties tools to detect the marks, and the detail of those tools is not yet published.[1]
Which question does a mark answer?
Two separate questions sit on this table: where the text goes, and who wrote it. The watermark makes the first traceable. I drew the same distinction on August 6 about the marking Suno announced: the mark manages a track's circulation once it leaves the platform and leaves open the question of what the track was made from. Anthropic's account now supplies, in the company's own words, the reason the second question stays open; if proofreading also leaves a mark, the presence of a mark does not settle authorship.[1], [3]
How much a detector is worth depends on the population it was evaluated on, where the threshold sits, and its false positive and false negative rates. None of those figures appear in the account. That absence limits what an outsider can check: the claim can be repeated but not tested. The strongest competing explanation is at hand too: the piece is a news report, and the company promises detailed technical documentation later, so the missing figures may show only that they have not been published yet.[1]
Where does the chain break when the weights are opened?
The same day, Meta published the weights of Muse Glimmer, a 30 billion parameter model, under the Apache 2.0 licence. The announcement describes the model as built for always-on agents that run on a Mac or a PC with a single consumer graphics card and no cloud connection, and says roughly 4 bit quantisation brings it under 20 gigabytes. The text covers training, quantisation and deployment in detail, and it does not mention content marking.[2]
The commitment Anthropic gave is applied at the model level and binds whoever operates the model. For a model such as Muse Glimmer, whose weights are downloaded and run on someone's own machine, no such operator exists, which turns the question of whether a mark survives distribution into one with two ends. Here is the testable part: if the promised detection documentation from Anthropic appears by October 31, 2026 and gives the evaluation population, the threshold and the error rates, how close the watermark comes to the authorship question becomes measurable. If that documentation does not appear, what remains is a tool that manages circulation, and the claim about identifying a source stays untested.[1], [2]