Anthropic Just Watermarked Claude. It Cannot Tell You Who Wrote the Essay.

The most useful thing about the Claude watermark Anthropic switched on this month is that it cannot tell you what your students wrote.

That sounds like a complaint. It is not. It is the first honest sentence anyone has been able to write about AI detection in three years, and it arrives in the exact week you are finalising a syllabus.

Here is what happened, in Anthropic’s own documentation. Embedded watermarks “will apply to all generated text,” and provenance metadata “will apply where Claude supports processing files,” following the C2PA open standard. The marking covers “output from supported models everywhere you use Claude, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag,” and it applies “wherever Claude is offered, worldwide.” Models released before August 2, 2026 are not covered yet, though the company says it is working on adding them. Detection details are promised in forthcoming technical documentation.

The trigger is regulatory. Article 50 of the EU AI Act, the transparency article, took effect on August 2, 2026. It requires providers of generative systems to mark synthetic output so a machine can detect it. Anthropic also signed the European Commission’s Code of Practice on transparency for AI generated content. So this is not a product idea somebody had. It is a compliance deadline with a date on it.

Now watch what happens next on your campus. Somebody forwards the headline to a department list. Somebody else replies that the detection problem is finally solved. And by the second week of September a colleague is sitting across from a student with a screenshot, and it is going to go badly, because the thing in that screenshot does not say what they think it says.

Three things the Claude watermark cannot tell you

Read the limitations the way you would read a methods section, because that is what they are.

It marks processing, not authorship. Read the sentence the company actually wrote: “Detecting a Claude mark tells you that the content may have been processed by Claude.” May. Processed. Not wrote. It does not distinguish a model that produced every sentence from one that fixed comma splices. A multilingual student who drafted in Portuguese and ran it through a translation carries the same mark as a student who typed one prompt and pasted the answer. Those are not the same act, and the signal does not separate them.

It dies in a rewrite. The documentation lists exactly when a mark will not be found: when the text “has been heavily edited, paraphrased, translated, or mixed into other writing,” when a file’s metadata “was stripped through format conversion, re-saving, screenshots, or other means,” or when “the passage is very short, leaving too little text for a reliable signal.” Sit with the incentive that creates. The student who pastes and submits gets caught. The student who runs it through a second tool does not. You have built a filter that selects for one more step of laundering.

Absence proves nothing. No mark does not mean a human wrote it. Older Claude models are not covered yet. Every other model on earth is a separate decision by a separate company. Unmarked work tells you the work is unmarked. That is the entire content of the finding.

The question the tool is answering is not the question you asked

Provenance marking is good infrastructure. Publishers need it. Newsrooms need it. Courts will need it. A signal that says this file passed through a model is genuinely useful when the question is where did this artifact come from.

Your question is different. You are not asking where a document came from. You are asking whether a specific person learned something. Provenance is about the object. Assessment is about the person. Those two questions can share a room for years without ever meeting, and the last three years of detection theatre have been one long attempt to answer the second question with an instrument built for the first.

This is the same shape as every other measurement failure in teaching. Seat time is not learning. Word count is not thinking. A provenance flag is not authorship. Each one is a real measurement of a real thing that happens to be adjacent to what you care about, which is exactly what makes it so easy to accept.

What to ask for instead

The move is not a better detector. It is a different request. Stop asking students to submit a finished object and start asking them to submit a trail.

Three checkpoints do most of the work, and none of them require software you do not already have.

Ask for the decision, not the draft. Before the work is written, students submit the choice they made and the option they rejected. Which framing, which source, which case. Two sentences. A model will happily generate a plausible pair, and it cannot generate the one your student defended out loud in a seminar on Tuesday.

Ask what changed. At the midpoint, one paragraph on what they now think is wrong with their own first version. Revision history is the cheapest authorship evidence in existence and almost nobody collects it.

Ask them to declare the tool. Not a confession box. A line item, the way you would cite a library database. What they used, for which step, and what they changed about the output. A student who names the tool and shows the edit has demonstrated more competence than a student who avoided it entirely, and that is not a concession. Article 4 of the same EU AI Act has since February 2025 required providers and deployers to ensure a sufficient level of AI literacy among the staff and others operating AI systems on their behalf. The declaration is the assignment where that happens.

Notice that none of the three asks whether AI was used. All three ask what the student did. The first question has an adversarial answer and a detection arms race behind it. The second has an answer only that student can give.

The scale problem, said out loud

This is where the honest version of the advice usually stops and the article ends with be intentional. So let us not do that.

Rebuilding one assignment this way takes about forty minutes if you know the course. That is a good trade. Then you look at your term and it is eleven more assignments, times three sections, and two of those sections are a preparation you have not taught since 2024. The method is not the hard part. The method is one page. The hard part is that a good decision made once has to be remade forty times before December, and it degrades a little each time you are tired.

That is the actual bottleneck in every faculty conversation about AI, and it is never the one on the agenda. It is the same gap between a fast tool and a slow institution that the three clock problem describes, reappearing at the scale of a single desk.

This is what SeedStacking is for. Seed is the first rebuilt assignment, the one you do by hand this week. Sprout is the checkpoint language you reuse without rewriting it. Grow is the term where the pattern holds across all three sections because the reasoning is written down instead of remembered. Harvest is a Course Record that carries your judgement into next year, so the version of you teaching this in January starts from your own decisions rather than a blank page. Do it once by hand and you own the method. That is your Teaching DNA, and it is the part no tool can supply, because it is made of choices only you have made.

The seed

The watermark is not the end of the AI integrity problem and it is not the beginning. It is a fact about a file. What it quietly does is take away the last reason to keep waiting for a machine to settle a question that was always going to be settled by the design of the work you assign.

The Claude watermark tells you a model was in the room. It will never tell you whether a student was.

Only your assignment can do that, and your assignment is the one thing in this entire argument you can rewrite before Monday.