The AI-generated watermark won’t save your child’s thinking

When the news came that ChatGPT’s answers would carry invisible watermarks in the EU, I could almost hear parents breathe out. Finally. Now the school will know.

I understand the relief. I don’t share it.

What actually changed

OpenAI has started adding an invisible signal to eligible ChatGPT and Codex answers in the European Union, in response to the EU AI Act. Anthropic, the company behind Claude, already does something similar for its supported models worldwide. The signal sits in the word choices themselves, a statistical pattern a detector can look for.

Now read the small print, because that is where the story is.

The detector isn’t public. For now only approved researchers and expert organisations can use it, not your child’s teacher. It works best on longer texts, and worse on maths, where there are fewer ways to phrase things. And it breaks easily: in OpenAI’s own tests, swapping a quarter of the words for synonyms dropped detection from about 92% to 17%.

Most important of all, the company says it plainly: “A watermark does not measure human contribution.” It can say a machine touched the text. It cannot say who did the thinking.

So who is going to catch them?

Do you really believe a signal that disappears after a few edits will stand between a twelve-year-old and an easy answer? That a teacher with thirty essays will run each one through a tool they’re not allowed to use? That fear of getting caught was ever what made a child learn?

No… we both know that fear never built a mind. At best it built a careful copier.

Thinking is connecting the dots

Here’s what worries me more than cheating. When a child asks AI to write the essay, the child doesn’t just skip the writing. They skip the part where the mind does its real work: [seeing that this idea connects to that one](future link: Blog #47 “Everything is connected”), that a chapter in history explains a headline today, that a question in science touches a question about life.

That act of connecting is learning. It’s slow, sometimes frustrating, and nobody can do it for you, not a teacher and not a machine. A child who hands the connecting over every time grows up with [lots of information and very little understanding]. They know many pieces and can’t see the picture.

What I’d ask the AI companies

If these companies can hide a signal in a machine’s word choices, they can do more. Imagine a small badge that doesn’t just say “AI touched this”, but shows how much of the work came from the human: the ideas, the structure, the edits. Seventy percent you, thirty percent the tool.

That badge wouldn’t be a trap. It would be a mirror. A child would see their own share grow from one task to the next, and a teacher would see the effort, not just the result. It would turn the question from “Did you cheat?” into “How much of this is yours?” That question teaches.

Today’s watermark can’t do this; OpenAI says so itself. But if the technology is already reaching into word choices, measuring the human part is the step that would actually serve learning. I hope they take it.

Why this matters to us

At Al-Andalusia, critical thinking is one of our core principles. It isn’t an extra skill added at the end of a lesson. It is the process itself: asking why, checking what you’re told, and seeing how one idea connects to another. A child who stops questioning stops learning, however many correct answers they hand in. That’s why every Al-Andalusia program is built on discussion and discovery, not on ready-made answers.

What you can do at home

You don’t need a detector. You need a few habits.

• Ask about the road, not the result. “How did you get there?” tells you more than any watermark. A child who did the thinking can walk you through it. A child who didn’t [goes quiet].
• Think first, then ask. Let your child write their own first answer, even a messy one, before any AI is opened. Then AI can question the draft, not replace it.
• Make AI the sparring partner, not the ghostwriter. “Find three weak points in my argument” builds a mind. “Write my argument” borrows one.
• Praise the messy draft. If only the polished result gets attention, children will reach for the tool that polishes fastest.
• One connection at dinner. Ask: what did you learn today that connects to something else you know? It’s a small question that trains the most important skill there is.

A firm line, and its honest limit

AI isn’t the enemy, and banning it doesn’t work. Children will [meet it everywhere], and they should learn to use it well. The line I’d hold is simple: AI may help you think; it may never think instead of you.

Where exactly that line falls will differ from child to child and task to task. That’s fine. What matters is that your child knows the line exists, and why.

The real reason

The watermark was never going to protect your child’s mind. Only your child can do that, and only if they understand what they lose when they give their thinking away.

They aren’t losing a grade. They’re losing the moment when everything suddenly makes sense.

So the question isn’t whether the school will find out. It’s whether your child will ever find out what their own mind can do. What will you ask them tonight?

Get the free Al-Andalusia module and our newsletter → [link]

Source
OpenAI joins Anthropic in text watermarking but limits ChatGPT rollout to EU – Emma Thompson, EdTech Innovation Hub, 8 October 2026