This week, Anthropic started weaving an invisible watermark into every piece of text Claude writes. It applies to cover letters, emails, code, company reports, anything the model touches. You will not see it, and neither will the person reading your application. But it is there, and under the EU AI Act it now has to be.
If you are applying for graduate jobs with any help from AI, this is worth five minutes, because the actual story is not the one in the headlines.
What the law actually says
Article 50 of the EU AI Act requires providers of AI systems to mark generated content so it can be told apart from something a person wrote. The transparency rules took effect on 2 August 2026, with a short extension for some existing systems running to December. Anthropic signed the EU's Code of Practice and rolled the watermark out worldwide rather than just to EU users.
For text, Anthropic describes it as a pattern in word choice that only a detector can read, invisible to anyone reading it normally. Google has done something similar with Gemini since 2024. OpenAI has said it will take a "layered approach" without giving specifics yet.
None of this is nothing. But read past the first paragraph of any of the coverage and the same caveats turn up every time.
What it does not actually do
The watermark survives copying and pasting. It does not reliably survive editing. Paraphrase a sentence, translate it, or blend it with your own writing, and the signal is gone. Anthropic has said as much itself.
Detection tools that would let an employer check for the mark do not exist yet. Anthropic says they are coming, with no date attached.
And even once they arrive, a positive hit will not mean what people assume it means. A watermark only shows that a model touched the text at some point. Ask Claude to fix the grammar on a paragraph you wrote yourself, and it comes back marked too. That is not the same as the letter being written for you, but a detector has no way to tell the difference, and the EU AI Act does not require it to.
This is a real transparency requirement with a real compliance deadline. It is also not going to be the thing that gets your application flagged, because it cannot be checked yet, and once it can, it still will not prove what a recruiter actually wants to know.
The thing that already gets you caught
A 2026 TopResume survey of over 800 hiring managers found that 67% believe they can spot an AI-generated cover letter. Their actual accuracy, tested, comes out closer to 52%. Recruiters overestimate their own detection, badly. But the method they are using is not a tool. It is pattern recognition, built up over hundreds of letters a season, aimed at exactly the kind of writing a generic prompt produces: "I am writing to express my strong interest," three adjectives in a row, a closing paragraph that tells you the opportunity would be relished.
We wrote about this problem last week, in a post about building tiero against Wikipedia's list of AI writing tells. A recruiter reading four hundred applications a season is not running your letter through a detector. They do not need to. The pattern does the work for them.
The watermark story gets this backwards for anyone applying to jobs right now. The risk was never that a machine would catch you. It was that a human would, because the writing had nothing in it that only you could have said.
Where tiero actually sits in this
tiero's Cover Letter Coach does not promise to get past a detector, watermark or otherwise, and we are not going to start now that there is a specific technology to point at. Detectors have a track record of being wrong in both directions, and a false accusation is a real cost to a real applicant. That is not a claim anyone can honestly make, so we do not make it.
What the Coach does is draw on your Experience Library rather than a template: the specific jobs, the specific things that went wrong and what you took from them, the detail a generic version of your CV does not hold. It gives you a full draft, and the FAQ on our own feature page says plainly that the point is you edit it, not that you accept it. It also runs the same AI-tells constraints in reverse, flagging the places a finished letter still reads as generated, including sentences you wrote yourself with no help from anything.
A hiring manager is not going to catch you because of what is hidden in the text. They will catch you because of what is missing from it. Bring the thing only you would know, and the watermark stops being relevant either way.
Written by
Adam Hannigan
Founder, tiero
Published
