On August 14, 2026, Anthropic published an explainer confirming that future Claude models will produce watermarked text. The company is doing it to comply with the EU AI Act, and it is not acting alone: around 190 organizations signed the EU's Code of Practice on Transparency of AI-Generated Content in July 2026, and the obligation to mark AI-generated text took effect on August 2.
If you write papers for school, that announcement is worth ten minutes of your attention — not because a watermark can march into your professor's office and accuse you of anything, but because the direction of travel is now obvious. Provenance is becoming checkable. Below is what the technology actually does, what it demonstrably cannot do, and the workflow that keeps you out of trouble either way.
What actually changed
Three facts do most of the work here.
First, it's a compliance move, not a crackdown. Anthropic states plainly that it is implementing watermarking to comply with the EU AI Act, and that other major model developers signed the same code of practice and will roll out their own watermarks. Nobody built this to catch students.
Second, it applies globally. Anthropic says it is applying watermarking worldwide at launch because it doesn't yet have a durable way to scope it by region. So "I'm not in Europe" is not a meaningful distinction.
Third, it's being backfilled. The EU law includes a transition period for models launched before August 2, 2026, and Anthropic says it is working to add watermarking to those older models over the coming months. Text you generate today from an older model may not carry a mark. Text you generate next semester probably will.
How a text watermark actually works
This is the part that surprises people. There is no hidden character, no invisible Unicode, no metadata tag, nothing appended to the end of your document. Nothing is added to the text at all.
A language model writes one word at a time, and at many points several words would work equally well. Anthropic's own example: "The weather today was cold and…" could reasonably continue with overcast or grey. The meaning barely changes. Normally that choice gets settled by a random number.
Watermarking swaps the source of that randomness. Instead of an arbitrary random number, the model uses a secret key plus the preceding few words to settle which word to pick. The output still reads naturally and the choices still vary — but anyone holding the key can check whether the sequence of choices matches the pattern the key would produce, and assign a probability that the model was involved.
Anthropic's implementation is a version of SynthID-Text, the approach Google DeepMind published in Nature in 2024. Anthropic reports no measured impact on quality, creativity, or readability, no extra tokens, and no change to price or speed. Critically, the watermark carries no identifying information — it cannot be traced back to you, your school, or your conversation. It speaks only to whether the model was involved.
Nothing is added to the text — no hidden character, no metadata tag, nothing appended. The watermark is the pattern in which word the model picks when several would work equally well, steered by a secret key.
What it proves — and what it doesn't
Anthropic is unusually candid about the limits, and those limits matter more to students than the headline does.
That last row is the one most coverage gets wrong. A watermark is only readable by someone holding the key that encodes it. Anthropic says it will "soon" offer a watermark detection API and is still working out the implementation details. Until such a service exists, no third-party checker can read a watermark. Anyone marketing watermark detection today is selling you something else.
A watermark hit cannot distinguish "the model wrote this" from "the model edited this." A clean result cannot show a paper was written by a person. Any policy built on either assumption will produce unfair outcomes — which is exactly why you want a paper trail of your own work regardless of what tools your school adopts.
Check your draft, not your luck.
TextChecker finds plagiarism, AI-generated text, grammar and spelling errors, punctuation slips, and citation problems — with a specific fix for each one.
Check my paperWhy the risk profile changed for students
Nothing about academic integrity policy changed on August 14. Submitting generated prose as your own work was already a violation at essentially every institution. What changed is the asymmetry.
Before, a student pasting AI output was gambling against detectors that guess from stylistic tells — tools that produce false accusations often enough that many universities distrust them. Now there's a second, structurally different signal in the mix, held by the model providers themselves and being backfilled into older models over time. You cannot check your own paper against it. Your institution eventually may be able to.
Two other details are easy to miss:
- Files carry provenance too. Anthropic says images and other supported files produced by Claude get a C2PA content credential in the file's metadata — an open industry standard already used by camera makers and photo editors. That's separate from watermarking and readable by any C2PA-aware tool. If you drop a model-generated chart or figure into an assignment, it may announce itself.
- Paraphrasing to hide is a losing trade. Light editing likely won't remove a watermark. A total rewrite will — but as Anthropic notes, at that point it's arguable whether the text is AI-generated at all. If you're rewriting every word anyway, you've done the work. You may as well have done it in the open and kept the credit.
A watermark check is probabilistic, not proof. It weakens on short passages, cannot tell text a model wrote from text a model edited, and can only be read by someone holding the provider's key.
If you use AI, cite it
The single biggest mistake students make isn't using AI. It's using AI silently. Disclosure converts a potential integrity violation into a documented methodology choice — assuming your course permits AI use in the first place, which is the thing to check before anything else.
Most major style guides treat a generative model as a software source. Formats shift between editions, so confirm against your assigned style manual and your syllabus, but the shape looks like this:
Two practical notes. Keep the prompt and the raw output — most instructors who permit AI want the transcript available on request, and a saved transcript is the cleanest possible answer to a challenge. And describe the use in a sentence in your methods or a footnote: what you asked for, what you kept, what you rewrote. "Used to generate a working outline, which I restructured" is a defensible position. Silence is not.
Citing AI does not override a course rule that prohibits AI. Policies vary sharply by institution, department, and instructor. If your syllabus is silent or ambiguous, ask your instructor in writing and keep the reply.
A workflow that holds up either way
The point of all this isn't to scare you off AI. It's that the workflow which survives scrutiny is also, unglamorously, the one that produces better papers.
- Start with real examples, not a blank prompt. Reading three completed papers on your topic teaches you structure, argument shape, and citation density faster than any generated outline.
- Build the argument yourself. Thesis, claim order, and what evidence supports what. This is the part being graded and the part you'll be asked about.
- Use AI where it's defensible. Brainstorming angles, stress-testing a thesis, explaining a concept you're stuck on, catching gaps in logic. Feedback on your draft leaves your words as your words.
- Keep your drafts and version history. Google Docs and Word both retain revision history. A visible progression from messy notes to finished draft is the strongest evidence a student can have, and it costs nothing to preserve.
- Run a check before you submit, not after. Source matching, citation formatting, and unsupported claims are all fixable at 11 p.m. the night before. They are not fixable after the deadline.
Disclosure is the move that protects you. Citing AI use turns a possible integrity violation into a documented methodology choice — provided your course permits AI at all, which is the thing to check first.
Where PaperDue fits
We've run a reference library for students since 1998, and our position on this hasn't moved: the examples exist to show you what a finished paper looks like, and the tools exist to make your own draft better. Neither is a substitute for writing it.
Our study guide library gives you completed papers across thousands of topics — structure, argument, and source use, all visible. Our writing tools handle the mechanical parts: citation formatting, outlining, titles. And TextChecker reviews a finished draft for plagiarism, AI-generated text, grammar and spelling errors, punctuation slips, and citation problems before you hand it in.
The current version of TextChecker does not read watermarks — as covered above, no third-party tool can, because detection requires the provider's key. That's changing. Anthropic has said it will release a watermark detection API, and we're building TextChecker to use it as soon as it's available.
When Anthropic's detection API ships, TextChecker will check your draft against it and tell you where Claude was likely involved in the text — before your instructor has the chance to. One scope note we'd rather state up front than bury: that check reads Claude's watermark specifically, not every AI model's. Each provider uses its own key, and some use a different method entirely.
The reason to want that in your own hands is simple. If a passage in your draft carries a mark, you have two good options — rewrite it in your own words, or disclose and cite it — and both take five minutes. Neither is available to you after you've submitted.
Frequently asked questions
Can my school tell if I used AI to write my essay?
Increasingly, yes — but not with certainty today. Anthropic began watermarking Claude's output in August 2026, and other major providers signed the same EU code of practice. A watermark check estimates the likelihood that a model was involved in a passage. It requires the provider's key, works better on longer passages, and cannot distinguish text a model wrote from text a model heavily edited.
Does editing or paraphrasing AI text remove the watermark?
Light editing generally does not. Replacing every word does — but at that point you've rewritten the passage yourself. Relying on paraphrasing to hide AI use takes roughly as long as writing the passage properly and still violates most integrity policies.
Does the watermark identify me personally?
No. Anthropic states the watermark carries no identifying information and can't be traced to a specific person, organization, or chat. It speaks only to whether the model was likely involved in producing the text.
Is it cheating to use AI if I cite it?
That depends on your institution's policy and your syllabus. Many courses permit AI for brainstorming, outlining, or feedback while prohibiting generated prose in a final draft. Disclosure doesn't override a rule that bans AI use — but undisclosed use is a violation almost everywhere.
Does PaperDue's TextChecker detect AI watermarks?
Not yet — and no third-party tool can today, because detection requires the model provider's key. Anthropic has said it will release a watermark detection API, and we'll integrate it into TextChecker when it ships, so you can check a draft for Claude's watermark yourself. That check will cover Claude specifically, not every AI model. Today, TextChecker reviews your paper for source matching, citation problems, and writing issues before you submit.
Sources
- Anthropic, "How Claude's text watermark works," August 14, 2026.
- Dathathri et al., "Scalable watermarking for identifying large language model outputs," Nature, 2024 (SynthID-Text).
- European Commission, Code of Practice on Transparency of AI-Generated Content, July 2026.
- C2PA — the content credentials standard for file provenance.
Check your draft, not your luck.
TextChecker finds plagiarism, AI-generated text, grammar and spelling errors, punctuation slips, and citation problems — with a specific fix for each one.
Check my paper
