Analysis and opinion · Research checked 7 September 2026
Imagine spending a weekend writing a personal essay. The experiences are yours. So are the argument, the awkward first draft and the ending you rewrote six times. Then you ask Claude to make the language flow better. If the finished version carries a detectable AI watermark, what should a reader conclude about your work?
That question has become concrete. Anthropic says text from its newly released Claude Fable 5.1 and Mythos 5.1 models carries watermarks. The marking applies worldwide across supported products, and copying the text into another document carries the watermark along. Older models are still being brought into the system. Claude’s marking guidance
The broader push is official, too. The European Commission says the AI Act’s marking obligations began applying on 2 August 2026, following a transparency code signed by companies including Anthropic, Google, Microsoft and OpenAI. European Commission announcement
The case for a detectable mark
For anyone worried about fabricated reviews or machine-written commentary pretending to be spontaneous human opinion, that sounds like progress. Readers deserve information about how something reached them. Publishers cannot sensibly promise transparency while treating every attempt to identify generated material as an insult to creativity.
And the technology is more subtle than a stamp across a page. Text watermarking can create a detectable pattern through choices between possible words during generation. The SynthID-Text research underpinning Claude’s approach tested feedback on nearly 20 million Gemini responses and found no statistically significant difference in positive or negative ratings between watermarked and unwatermarked outputs. That supports the case that marking can preserve the reading experience. SynthID-Text study in Nature
A watermark cannot explain the whole collaboration
The awkward part begins when identifying a tool’s involvement becomes a judgment about the person using it.
Anthropic explicitly says its watermark cannot distinguish text Claude wrote from text it heavily edited. Translations carry a watermark because Claude chooses the translated words. Light proofreading may produce too few changes for its involvement to register. The mark also contains no information identifying an individual user or their conversation. Anthropic’s watermark explanation
Those distinctions matter. Consider someone translating an account of their own childhood into English. The wording involves AI; the memory does not. Or a researcher who supplies the evidence and argument but asks for clearer sentences. A single label could leave readers with a very different impression from a description of the actual assistance.
There is a fair objection here: heavy rewriting really does contribute to a piece. Choosing the rhythm, tone and structure is part of writing. Calling every substantial intervention “polishing” can conceal how much work the system performed. Someone who generates an entire essay and changes its opening sentence should not expect the same description as someone who corrects a few commas.
The difficulty is deciding where those cases belong. A watermark does not settle that debate. Anthropic says detection estimates whether Claude was involved; it cannot establish that an unmarked passage was human-written, and short samples are harder to assess. The underlying research also describes detection trade-offs involving text length and the amount of choice available to the model. Anthropic’s limits, the research
Who gets to interpret the result?
There is another practical wrinkle. As of 7 September, Anthropic’s detection service is in private preview for eligible organisations, including educational institutions, media and researchers. Wider access is planned. That leaves a question worth asking before detection becomes routine: how can writers understand or challenge conclusions drawn using a service they may not be able to use themselves? Current detection access
Our view is that watermarks are worth developing, and institutions should explain what they mean before relying on them. A useful disclosure would describe the contribution: translated, substantially rewritten or generated from a prompt. Treating those as interchangeable would throw away precisely the context transparency is supposed to provide.
Writers should also be honest about substantial assistance. Being human does not make every word personally crafted, and having an original idea does not erase a machine’s contribution to expressing it. But readers deserve enough detail to make that judgment themselves.
If you wrote the ideas and first draft, then let AI substantially rewrite the language, what description would feel fair beside your name—and would you accept the same description on someone else’s work?
Leave a Reply