From this month, every sentence Claude writes carries a mark you cannot see
On 11 August Anthropic confirmed that models released after 2 August embed an invisible watermark in their text output. Copy it into Word, into an email, into a piece of homework, the mark travels with it. The EU AI Act's transparency code took effect on 2 August, but this isn't EU-only: it applies worldwide, across their products. Google, Meta, Microsoft and OpenAI have signed up to the same code.
Teachers have wanted this for two years. So it's worth being precise about what it actually detects.
📄 Picture one assignment and three pupils
A had AI write the whole thing.
B wrote it themselves, then asked AI to fix the grammar.
C wrote it themselves and never touched AI.
A and B produce an identical result: marked. C isn't.
And there's a fourth. D wrote none of it, but used a different tool, or an older model, or simply retyped the AI's text by hand, also unmarked.
So the mark answers "did a machine touch this text". It does not answer "whose thinking is this".
Anthropic's own documentation is refreshingly blunt: a detected mark "provides a signal that content was processed by Claude, but is not fully conclusive", and a lack of a detected mark "doesn't mean the content wasn't AI-generated or processed".
D wrote none of it, but used a different tool, or an older model, or simply retyped the AI's text by hand, also unmarked.
🔍 The part almost no coverage mentions
Right now you can't check anyway. There is no public detector. The official line is that they're working to let users and third parties detect it, with technical documentation to follow. So what your child's school can do with this in practice this week is nothing. This is groundwork, not a starting whistle.
🎓 The bit I think is worth sitting with
A school rule that says "don't use AI" now has half a way of being enforced. But what it enforces is whether a machine was involved, not whether the pupil thought. The second one is the thing we actually want to teach.
That gap forces an awkward incentive. If the penalty attaches to the mark, then the rational move for a pupil is to stop using AI for grammar help, because tidying your own sentences and having the whole thing ghostwritten look the same to the mark. Is that the lesson we want them to draw?
I use AI daily in my own work, so to be fair: I don't think this is the wrong move. Provenance that survives copy-paste is a good thing long term, and the regulation requires it. I just want the distinction kept straight, it's a processing record, not a cheating verdict.
If you ran a school, would you use the mark to catch people, or simply ask pupils to declare which parts they used AI for, and how?
https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
