IGNITE AI Detection
AI Text Watermarks: What They Can and Can't Tell an English Teacher
Since August 2026, a European law has required AI companies serving the EU to mark what their tools write, and more of them now hide a watermark in the text itself. If you've seen the headlines, you may be wondering whether AI detection finally works. For a Grades 7 to 12 English teacher, the short answer is no, and the reasons are worth knowing before the question comes up at a staff meeting.
What Changed
A new European law requires AI providers serving the EU market to mark content their systems generate. Around 190 organizations, including Anthropic, Google, and OpenAI, have signed the code of practice that sets out how. Anthropic, which makes Claude, explained its approach in August 2026, and applies it worldwide, not only in Europe. Google has watermarked text from its Gemini app since 2024.
I use Claude when I build IGNITE resources, as the AI disclosure page explains, so this affects my own work too. What follows is my own reading of what Anthropic has published, and it doesn't speak for Anthropic.
How a Text Watermark Works
A chatbot writes by choosing a likely next word, over and over. Often several words are about equally good: a cold day might be overcast or grey, and the meaning barely changes. Normally that choice is settled at random. A watermark settles it with a secret key instead. The text reads exactly the same, but anyone holding the key can check whether the pattern of choices matches. Nothing is added to the text, and there are no hidden characters to find.
What It Can't Tell You
This is the part that matters in a classroom. By Anthropic's own account, its key can answer only one question: "What is the likelihood this was partly written by Claude?"
- 1It can't confirm a student wrote somethingA missing mark doesn't prove human writing, and it can't detect another company's AI, because each company uses its own key.
- 2It struggles with short writingShort passages offer fewer word choices to carry the pattern, and most student answers, paragraphs, and exit tickets are short.
- 3It can't tell writing from editingIt can't distinguish Claude writing a piece from Claude heavily editing one, light proofreading may leave too little to detect, and a complete rewrite removes it.
Even when a mark is found, Anthropic describes it as a signal that content may have been processed by Claude, not proof.
You Probably Can't Check Anyway
Detection is in private preview. It's open to organizations eligible under EU law, such as regulators, media, researchers, and educational organizations, and to companies with their own obligations under the law. Anthropic says it plans to widen access over time. For now, most classroom teachers have no way to check a piece of writing for a watermark at all.
The Fairness Problem Hasn't Gone Away
Watermarks avoid one flaw of detection software: they don't guess from writing style, so a student who writes formally, or in a second language, isn't flagged for sounding like AI. But they raise a new question. A translation carries the watermark in full, because every word was chosen by the AI. A multilingual student who drafts in their first language and translates with an AI tool will carry the mark, even though every idea is their own. Whether that's acceptable is a question for your school's policy, and a watermark can't answer it.
What This Means for Your Assignments
The useful question hasn't changed. It isn't "did a machine touch this?" It's "can this student show their own thinking in front of me?" A watermark, at best, answers the first. Dated process work, writing done in class, and an honest declaration of what AI did answer the second, and they work whichever AI a student used. That's the case made in the piece on detectors, and it still holds.
A clear AI use declaration also covers the grey zone a watermark can't see: the difference between a tool that fixed three commas and one that wrote the argument. Asking students to say what the AI did, what they did themselves, and how they checked the result is more useful than any mark, because it teaches them something.
Where to Start
With one assignment you already use. The Audit asks eight questions about it and shows how much of the thinking a chatbot could do for your students, watermark or not, and the first thing worth changing.
Put One Assignment Through the Audit
The free Audit takes an ordinary task through eight questions and names the one thing worth changing first. Five minutes.
Confirm by clicking the link in the email that follows, and the files arrive straight after. An occasional email after that, usually every week or two. Unsubscribe in one click. A personal address works better than a board or district one.
Questions Teachers Ask
Does a watermark prove a student cheated?
No. At most it suggests that one company's AI was involved at some point. It can't show who did the thinking or tell writing from heavy editing, and Anthropic describes a detected mark as a signal, not proof.
Can I check student work for a watermark?
In most cases, not yet. Detection is in private preview for organizations eligible under EU law. Anthropic says access will widen over time, so this answer may change.
Does Anthropic endorse the IGNITE Framework?
No. This article is my own reading of Anthropic's published explanation, and it doesn't speak for Anthropic. I use Claude to help build IGNITE resources, and the AI disclosure page explains how.
Where This Comes From
Everything above is drawn from Anthropic's own public explanations. Read them in full rather than relying on any summary, including this one.
- 1Anthropic, How Claude's text watermarking worksAugust 2026, updated September 2026. Available at anthropic.com.
- 2Claude Help Center, How Claude marks AI-generated contentAvailable at support.claude.com.
Developed by George O'Toole: over 35 years in secondary English, Head of Department, Educational Technology Coach, Faculty of Education instructor. Meet George