Invisible AI text watermarks: what they mean for a translated book
Anthropic, the company behind the Claude models, said this month that it embeds an invisible machine-readable mark in text its models produce. For an author publishing a translated book this changes less than it sounds, but it is worth understanding exactly what an AI text watermark says. Mainly because it does not say what most headlines implied.
What was announced, and when
TechCrunch reported on 11 August 2026 that the company would mark text produced by its models, and its official help page sets out the details.
According to that page, models launched from 2 August 2026 onward support the marking from launch, and earlier models are being brought in. The marking works in two ways: text carries an imperceptible mark woven into the text itself, and files such as images carry signed provenance information.
The background is regulatory. According to that report, the European Union's transparency rules took effect on 2 August 2026 and require content generated with artificial intelligence tools to be marked in a way other systems can identify. The marking was applied worldwide, not only in Europe.
What the mark says, and what it does not
This is the important part, and it is stated openly on the company's own page.
The mark indicates that the content may have been processed by the model. It does not indicate who the author is, and the company states explicitly that the model may not be the original source of the text. In other words, it marks contact — not ownership and not originality.
And one more point worth reading twice: according to that page, the mark can be obscured when text goes through heavy editing, paraphrasing or translation. The TechCrunch reporter also noted that it is not clear how much editing is needed to remove it.
There is an irony there that touches our readers directly: translation is precisely one of the operations the company lists as capable of erasing the mark.
What this changes for a translated book
In practice, less than it seems. Three things are worth separating:
This is not a store rule. No store has announced that it checks files for watermarks. What Amazon does ask for is your declaration about tool use, and we wrote about the difference between a translation produced by a tool and editing assisted by one in what Amazon requires you to declare.
This is not a quality test. Marked text can be excellent and unmarked text can be poor. The checks that actually find problems in a translated book are entirely different, and we set them out in how to check a translation in a language you cannot read.
It is a sign of direction. Even if the marking itself is loose, it shows where the wind is blowing: transparency about tool use is becoming the default rather than a question asked only when somebody complains.
Why it is still worth paying attention
Because this trend touches a sore spot in the book world right now. The Christian Science Monitor reported in August 2026 on a run of books pulled from sale or dropped by their publishers after suspicion that they were written with artificial intelligence tools — in the cases it describes the authors denied it, and no proof was produced.
That same report notes that the detection tools themselves are unreliable, and that a 2023 study found they were biased against writers whose first language is not English. In other words: the suspicion alone was enough, and the tool meant to settle it does not really settle anything.
The distinction worth holding on to is between writing and translation. A book written by a person and translated with a tool is not the same thing as a book written by a tool, and store rules do distinguish between the two. Public discussion does not always distinguish, which is why it pays to describe your own process accurately.
What to actually do
Three things, all of which take under an hour:
- Answer the disclosure question correctly when you upload the book. That is the one field that genuinely gets checked.
- Keep a short record for yourself: what you wrote, at which stage a tool was used, who reviewed the result. One page is enough, and it is worth a great deal on the day somebody asks.
- Do not try to strip marks. That is not what gets checked, and it is not what protects you.
And if you are still at the stage of deciding how to translate the book at all, the fair comparison between the options is in AI translation versus a human translator, and the remaining steps are laid out in frequently asked questions.
Frequently asked questions
What was actually announced?
Anthropic said in August 2026 that it marks text produced by its models. According to its own help page, models launched from 2 August 2026 onward support the marking from launch, and older models are in the process of getting it.
Can you see the mark?
Not with your eyes. It is an imperceptible mark woven into the text itself, not something that shows on screen. According to the company it travels with the text when it is copied and pasted elsewhere.
If I translated a book with AI, does it carry a mark?
Possibly. But the company itself notes that heavy editing, paraphrasing or translation can obscure the mark, so this is not a definitive way to establish where a text came from.
Does this mean my book will be rejected by a store?
No. At most the mark shows that text passed through a model, not that the book is disallowed. Store rules are about declaring your use of tools, which is a separate matter entirely.
So what should I actually do?
Answer the disclosure question correctly when you upload the book, and keep a short record of how you worked. Those two are worth more than any attempt to guess how detection technology will behave.