What AI watermarks actually mean if you run a business

If you run a business that publishes anything, whether a website, a newsletter, product descriptions or support articles, you have probably seen a headline about AI watermarking and wondered whether it is your problem. Here is the practical version, without the alarm.

What changed

Article 50 of the EU AI Act became enforceable on 2 August 2026. It requires providers of generative AI to mark their outputs in a machine-readable way. Anthropic responded by watermarking text from Claude models released on or after that date, with older models to follow during the transition period the law allows. The obligation in that article sits on the provider of the model, not on you for using it.

So the immediate answer for most businesses is that nothing is required of you today by this particular change. That is not the same as nothing being worth doing.

What it means for text you have already published

• Anything drafted with a covered model may carry an invisible mark. It does not change how the text looks, reads or ranks.
• Light editing does not reliably remove it. A full rewrite, where effectively every word changes, does.
• Translation tends to keep it, because the translated words were also chosen by the model.
• Short pieces carry little signal. So do dense factual passages, where there are few alternative words to choose between.
• Images are handled differently: rather than a watermark, generated image files carry C2PA content credentials, which live in file metadata and can be stripped by ordinary processing.

Note the last two together. The mark is weakest exactly where content tends to be shortest and most factual, and image provenance survives only as long as nobody re-saves the file. Anyone selling you certainty about either is overselling.

The one thing actually worth doing

Write down what your business does, and tell people. Not because a regulator is coming for your blog, but because the disclosure is going to be worth more than the detection.

In two years the interesting question will not be whether you used AI. It will be whether you said so before anyone asked.

A policy can be three lines. Where AI assists drafting. Where a named person reviews before publication. What you will never let it do. For most businesses that list should include anything with a number in it, anything making a claim about a customer, and anything with legal effect.

What not to do

Do not buy a detector and start pointing it at your team, your freelancers or your applicants. Detection has one honest direction: finding a mark is evidence of something, and failing to find one is evidence of nothing. A person who wrote every word themselves produces exactly the same result as a person the tool simply cannot identify, and you will not be able to tell them apart.

And do not rewrite good published work purely to strip a mark. It is effort spent making your content worse in order to hide something that is not, in most contexts, a problem.

Where this is heading

Provenance is quietly becoming infrastructure. Software has been signed for years so that an operating system can name the company behind it. Photographs are gaining content credentials. Text is now marked. The direction is consistent: publishing is drifting toward a world where things carry a record of where they came from.

Businesses that already behave as though that were true will find the transition uneventful. That is the whole strategy, and it costs nothing to start now.