Claude is watermarking its writing. What that means for your business

On 14 August 2026, Anthropic published an explainer saying future Claude models will produce text carrying an invisible watermark. In plain terms: the words Claude writes for you now contain a hidden statistical pattern that Anthropic can later test for, and every other major Ai provider serving the EU has agreed to do something similar.

The part most coverage skips is how weak the signal is by design. A watermark check does not say who wrote a document, or whether it is any good. It says one thing: Claude was probably involved somewhere. Anthropic is unusually blunt about those limits, and they are the whole story for anyone running a business on Ai-assisted writing.

What actually changed?

Ai models write one word at a time, choosing from a shortlist of sensible options. Where two options are equally good, the choice is settled at random. Watermarking swaps the source of that randomness for a secret key, so the finished text carries a faint pattern that anyone holding the key can test for. Nothing is added to the text and there are no hidden characters. Anthropic uses a version of the SynthID-Text method published by Google DeepMind in a Nature paper in 2024.

The reason is regulatory, not commercial. Anthropic says it signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside several other major Ai providers and around 190 signatories in total, and that from 2 August 2026 the EU requires Ai providers serving its market to mark Ai-generated content. It is applying the watermark globally at launch because it says it has no durable way to limit it to one region yet. Older Claude models sit inside an EU transition period and will have watermarking added over the coming months. The announcement does not describe a way to switch it off.

Source: Anthropic, How Claude’s text watermark works

Will anyone notice a difference in the writing?

No. Anthropic says a watermarked response is indistinguishable from an unwatermarked one to a reader, and that internal testing showed no impact on content, creativity or readability. It points to the SynthID-Text paper, where Google DeepMind served a watermarked model to part of its Gemini traffic and found no statistically significant difference in user ratings against the unwatermarked model.

Does it cost more or slow things down?

No on both counts. In Anthropic’s own words: “Watermarking doesn’t require extra tokens, and will not be more expensive.” The effect on speed is described as negligible, because no additional text is produced.

Can a watermark be traced back to me or my business?

No. Anthropic says the watermark carries no identifying information and cannot be traced to a specific person, organisation or chat. A test tells you Claude was likely involved. It cannot tell you which account, which company or which conversation produced the words.

Can anyone actually check a document today?

Not yet. Anthropic says it “will soon be offering a watermark detection API” and is still working out the details. Until that exists, the watermark is being written into text that almost nobody can read. If you are worried about a client or a procurement team testing your work, right now they cannot.

What about images and files?

Files are handled differently. When Claude produces a supported file type such as a .png, .jpg or .svg, Anthropic attaches a content credential to the file’s metadata using C2PA, the open standard camera makers and photo editors already use to record where an image came from. That is a label sitting alongside the file rather than a pattern hidden inside it, so any C2PA-aware tool can read it, and it lasts only as long as the metadata does.

What if Claude only tidies up my writing?

Then there is likely to be nothing to find. The watermark only attaches to words Claude chooses, so on a light proofread, where nearly all the words are the person’s own, Anthropic says the changes may be too few to register. The same goes for short passages, strictly factual sentences where only one word is correct, and largely for code. Translation is the opposite case: Claude picks every word, so a translation carries a full watermark.

What to ignore

Ignore the idea that this makes Ai writing “detectable” in the way people fear. Anthropic states the limit plainly: “A watermark can only determine that Claude was likely involved with the content at some point.” It cannot separate writing from heavy editing, cannot confirm anything was written by a human, and cannot recognise text from a different Ai, which would use a different key or method entirely.

Ignore the panic about existing Ai detector tools too, which is a separate and much shakier thing. Those services have no key. They guess from stylistic tells, and Anthropic names two of them: models are fond of the construction “this isn’t [X], it’s [Y]”, and use the word “quietly” far more than you would expect. Guessing from phrasing is not the same as testing for a watermark, and it is why detector tools produce false accusations.

What should a small business do next?

Two small things, both free. First, keep your own record of the work behind anything you publish: the brief, the source figures, the version history in Google Docs or Word. Provenance you control settles a client dispute in a way no detector result can. Second, spend ten minutes deciding which jobs Ai drafts and which ones a person writes from scratch. Client apologies and anything a regulator might read are usually worth writing yourself, then using Ai to sharpen. That applies equally to everyday tools like Gmail’s Help me write and to a full assistant such as Claude Opus 5. If you want it written down properly rather than held in your head, that is the kind of job a Breezi membership hour covers.