A Watermark Can Prove Processing.

It Can’t Prove Authorship.

Anthropic began embedding an invisible mark in Claude-generated text this month: a signal, the company says, that content “may have been processed” by its models, not proof of who wrote it. The mark applies to a paragraph you drafted yourself and asked Claude to tighten, exactly as it applies to a paragraph Claude wrote from nothing. Anthropic itself draws no distinction between the two, and that nuance is unlikely to survive contact with whoever, or whatever, reads the signal next.

By William Douglas Baker

Founding Editor

Acomma gets fixed. A paragraph gets marked. That’s not a hypothetical — it’s Anthropic’s own documented scope for the mark: proofreading, translating, summarizing, and converting human-written work all qualify, the same as text a model wrote from a blank page. Every Claude model launched since August 2, 2026 carries it from first release; Anthropic says it’s working to extend it to older models as well. The mark travels through copy and paste. No opt-out is documented.

Anthropic is careful about what the mark claims. It never says Claude wrote the words — only that Claude touched them somewhere along the way. The thinking inside them can still belong entirely to the person who typed them. An absent signal proves the opposite even less: heavy editing, a short passage, a format conversion, or simply an older model can all leave no detectable trace at all.

What the mark can carry is content provenance — evidence that a set of tokens passed through a specific model at some point in their production history. What most people will assume it carries is intellectual provenance — evidence of who originated the idea, the research, the judgment, the sentence. Those are not the same claim, and a mark woven into the words themselves was never capable of making the second one. It can’t distinguish a writer who dictated every word and asked for a punctuation pass from a writer who supplied a topic and walked away. Digital Derangement Syndrome™ already names this pattern at the answer-engine layer: search and citation misclassifying real authority. This is the same failure one layer down, inside the AI’s own disclosure about itself — Trust Transfer Failure, a signal that exists but can’t bear the trust it’s being asked to carry.

What the mark proves

What this doesn’t fix

A detection layer this narrow won’t resolve what it’s being asked to resolve. Employers, publishers, platforms, and eventually other AI systems will encounter this signal and read it as a verdict on authorship, because that’s the question everyone wants answered. The mark can’t answer it. The leap runs in three short steps: a machine touched this, a machine wrote this, the person who signed it is lying. Each step requires exactly the kind of judgment no watermark was built to carry. And once classifiers and answer engines start consuming these signals automatically, the question stops being whether a human misreads the mark. It becomes whether a machine does, at scale, with no one positioned to correct it.

None of this makes the watermark dishonest. Anthropic’s own documentation is more careful about the mark’s limits than most of the coverage of it has been. But notice the question the mark never claims to answer: whether the words are true. A marked paragraph can be accurate. An unmarked one can be false from its first word, and confident falsehood never needed a machine’s help. Provenance is not a warranty. What the mark leaves unbuilt is the harder problem: a durable record of the judgment that made the words true.