Do you know if the tool your team drafts proposals with marks its text? If you don’t, you’re in good company, and I think it’s worth finding out this week. One of the major AI writing models now marks its text as it writes. The mark is invisible on the page, readable by a machine, and it lives in the word choices themselves. If your team drafted a proposal with that tool, the page may already carry AI content labelling, and nobody in your firm made that call.
The rule behind it is Article 50 of the EU AI Act, in force from 2 August 2026 with a grace period to 2 December 2026 for systems already on the market, which asks AI providers to mark generated text in a machine-readable way. A Calgary firm using those tools inherits the mark even if no Canadian rule asks for it.
No tag travels with the file, and no code hides under the text. The mark is a pattern in how the words were picked, a lean toward some words and away from others that software can measure. Copy the text into a fresh document and the pattern comes with it, because the pattern is the text. Not every writing tool does this. Yours might. And most owners have never asked.
Where does the mark actually live?
In the choices the model made while it was writing. Each next word gets picked from a field of likely words, and the marking tilts those picks, gently and on purpose, so the finished text leans in a direction a checker can measure.
Nothing shows on the page. No watermark image, no footer, no hidden line somebody can delete. The words look ordinary because they are ordinary words, chosen with a thumb on the scale.
Think of the label sewn into a shirt. You never see it while the shirt is being worn, and it changes nothing about how the shirt fits. It is simply there, part of how the thing was made, waiting for anyone who knows where to look.
The label is not attached to your words. It is woven out of them.
Why is this arriving now?
Because European rules on labelling machine-made text pushed one maker to build the mark in at the source. If text written by a machine has to be identifiable, the cleanest place to answer is the moment the text is produced, so the mark goes on before the words ever reach your desk.
A way for anyone to check a page of text is on the way. When it lands, the question stops being technical. A client can paste your proposal into a checker, and they will not phone first to warn you they did.
What should your firm’s line be?
Three decisions, made once and written down. What gets disclosed. Where the note sits. And which person signs the finished piece. I’d make all three before a client asks.
Start with the work your name rides on. Proposals, reports, letters, anything a client reads as yours. Decide if those pieces carry a plain note that AI helped draft them, and keep the note in the same place every time. Internal drafts and working notes sit on a different shelf; nobody labels a brainstorm. Then name the person who reads the final piece before it leaves and signs it, because that signature is the part no tool can lend you. A client who finds your note learns you run your tools in the open. A client who finds the mark first learns something else entirely.
I argued the same habit for pricing tools, to write your rules first while you are calm, because calm rules hold up better than improvised ones. And keep this line the way AI rules work best in a firm, as a habit, not a binder: three sentences the team uses every week, not a document opened once a year.
None of this asks you to stop writing with AI. The drafting help is real, and the firms you compete with use it too. The line is about who decides what your clients get told, and the answer should be you, in advance, in writing.
Which writing tools should you check this week?
Make the list first. Every tool anyone in the firm writes with, including the writing help built into your email and your documents. Then ask each maker, in writing, one question. Does your text output carry a mark, and can it be checked.
Some will answer clearly, and the vague answers tell you something too. Then write your three sentences. What gets disclosed, where the note sits, who signs. Put them where the writing happens, inside the template your team opens every day. The line you write this week is the one you will be glad exists the first time a client asks. Setting that line, and the habits around it, sits inside the AI Strategy Roadmap.
Short answers on AI content labelling
Is AI-generated text being watermarked?
Some of it. Under Article 50 of the EU AI Act, providers of generative AI must mark synthetic output, text included, in a machine-readable way, with a grace period to December 2, 2026 for systems already on the market. Some marks live in the word choices, so copied text carries them along.
Should my business disclose when AI helped write client work?
Decide it once and write it down. For proposals, reports and letters clients read as yours, choose if a plain note says AI helped draft them, keep it in the same place every time, and name the person who signs the final piece. Internal drafts need no label.
How do I find out if my AI tool marks its text?
List every tool your firm writes with, including the writing help built into email and document apps. Then ask each maker in writing if its text output carries a mark and if that mark can be checked. A vague answer tells you something too.
Send each tool maker the one question in writing this week. More of these questions land Friday mornings in the AI Newsletter with ZAK.
4 Comments