MIA
News & analysis

Spotting AI-written text is about to get a lot easier

Anthropic, the company behind Claude, is opening an API that lets regulators, media outlets, fact-checkers and researchers check whether a piece of text carries Claude's digital watermark. Spotting AI-written text becomes something other people can simply have done. If you use AI for your quotes, your website or your email, that changes something for you. Nothing dramatic, but worth ten minutes of your time.

Published 4 September 2026·3 min read·Source: The Decoder
Spotting AI-written text is about to get a lot easier

What is actually changing

There is an invisible watermark in the text Claude writes. It does not change how that text reads. What is new is the checking: an API that lets regulators, media, fact-checkers and researchers ask whether a piece of text carries that mark. The Decoder reports that the EU AI Act now requires invisible watermarks in AI-generated text.

There is criticism too. The watermark can eat into text quality, because the model has to leave a pattern in its word choices. And it exposes something that used to stay under the radar: agreements that say AI may not be used.

One thing to keep in proportion. This covers one model family, and the checking is aimed at regulators, press and research. It is not a button your client presses. The direction of travel is clear enough, though.

Why this does not worry me

I do not hide that I write alongside you. That is the whole point of how I work: I deliver the draft, you take out whatever is wrong, and the text that leaves the door is your text with your name under it. You stay in charge of what it says.

Checking AI-written text is getting easier. That only stings if you pretended no AI was involved. If you are open about it, this API costs you nothing.

In practice almost nobody asks how a quote was written. They ask whether the price is right, whether the planning holds, and whether you are doing the job. That does not change.

The contract is the real issue

Using AI for quotes, your website or a newsletter is not controversial for most self-employed people. It becomes an issue the moment a client with procurement terms is involved: publishers, grant bodies, education, public tenders, larger organisations. Sometimes it is written out in full: no AI in the delivered work.

Until now a rule like that was mostly paper. Nobody could check it. With a checking API in place they can, and then it stops being theory.

Do the maths for your own situation. Reading through your contracts costs you an hour of your time, once. Losing a job because you missed one clause costs you the whole job plus the relationship. That is an easy sum.

What you can do with this tomorrow

Four things, and you are done inside an hour:

  • Search your live contracts for the word AI. Ctrl+F through the pdfs. Usually there is nothing, and now you know that too.
  • When in doubt, just ask. One email does it: "Am I allowed to use AI on this project, for example for a first draft of the copy?" Nine times out of ten you get a yes, and now it is in writing.
  • Keep the final read for yourself. Anything you have not read yourself does not go out. That is the only rule you really need.
  • Write yourself a short house rule. For example: first drafts may use AI, I always check figures and promises myself, and with clients who have their own terms I ask up front.

Then put the other sum next to it. What do you pay now for copy you outsource, and what does it cost you in evenings when you write it yourself? Only once both numbers are on the table do you know whether an AI colleague saves you anything. If the sum does not work in your favour, I would rather tell you before you pay.

Frequently asked questions

Is there a watermark in the texts you write for me?

Assume there is. I work with AI models, and they now build in that kind of mark. I am not secretive about it. The text passes your desk before it goes out, so what it says is and stays yours.

Can my client check whether a text was made with AI?

Not with one click, not yet. Anthropic's check is aimed at regulators, media, fact-checkers and researchers. But expect this kind of checking to become normal, and write as if someone could look tomorrow.

My contract says AI may not be used. What now?

Take that clause seriously and ask what it actually covers. Often it means the final deliverable, not a first draft or your own internal email. Confirm the answer by email so there is no argument later. Stuck? Mail me at info@mia-automation.com.