An AI-written article needs no label. Your proofread email gets one

Start with what the mark actually is, because it is not a stamp on the page.

When a model writes, it picks each word from a set of reasonable options. Anthropic uses a version of SynthID-Text, the method Google DeepMind published in Nature in 2024. It draws several candidate words from the model’s own distribution, then runs them through a knockout bracket. A secret key and the preceding words decide each round.

The vocabulary is never split, and the model is never pushed towards words it would not otherwise pick. Anthropic says so directly, and the paper reports no measurable quality difference across roughly 20 million Gemini responses. What changes is only the source of the randomness.

One word proves nothing. Across 500 or 1,000 words, a detector holding the key can see the pattern.

Anthropic began marking Claude output on 2 August, worldwide rather than only in Europe. It published the mechanism on 14 August, in a post that is unusually specific about where the mark fails. It still has not released a detector.

It marks what Claude touched, not what Claude wrote

This is the part most coverage skipped, and Anthropic states it plainly in its own support article.

A detected mark means the content “may have been processed by Claude”. It does not mean Claude wrote it. Ask the model to translate or summarise your own writing, and the output carries the mark anyway, because Claude chooses every word. Anthropic confirms that for translation in as many words.

Proofreading is the exception, and Anthropic drew that line on 14 August. When Claude edits a person’s text, nearly all the words are still the person’s, so there is “very little (if anything)” for the mark to attach to. A light grammar pass may leave nothing detectable at all.

The reverse also holds. No mark does not mean no AI. Short passages, heavy paraphrasing, older models and stripped file metadata all produce clean text that a machine still helped make.

Anthropic names the same weak spots itself. Detection performs poorly on short passages, because there are too few choices to test. It thins out on factual writing, where accuracy leaves only one right answer, and on code, which has to be exact to run.

So the signal has two failure directions at once, and the company says so upfront.

It also proves less than a reader might assume. The mark answers one question, which is how likely it is that Claude was involved. It cannot confirm a human wrote something, it cannot recognise another company’s AI, and it cannot tell “Claude wrote this” apart from “Claude heavily edited this”.

The law exempted the grammar fix, and so did the physics

Here is where it gets strange. The EU AI Act does not require marking when a system performs an assistive function for standard editing, and the Commission’s own example of that is grammar correction.

Anthropic marks at the model level regardless. Ars Technica called the approach “nuke it from orbit”.

The reason is structural rather than ideological. A watermark applied at the model level cannot distinguish a full draft from a comma. It marks the output, and the output is all it sees.

What saves the grammar fix is not the policy but the maths. A comma gives the watermark almost nothing to work with, so the exempted case is also the case where the mark is weakest. Brussels and the algorithm arrived at the same answer for entirely different reasons.

And the label almost nobody has to show

Now look at the other half of Article 50, the part that governs what publishers must tell readers. It is far narrower than most people assume.

An AI-written novel needs no label. AI-generated marketing copy needs no label. Text that informs the public on matters of public interest does need one, unless a named and accountable human editor has reviewed it.

Put the two halves together and you get the contradiction. Everything Claude writes carries a mark at the model level, while a fully synthetic article can reach readers unlabelled because an editor looked at it.

The EU has been building this machinery for a while. It has already made compulsory labels for synthetic content, and it has given itself powers to inspect and fine models directly. The penalty here runs to €15m or 3% of worldwide annual turnover.

The objections arrived within a day

Investor Bill Gurley argued that if only Anthropic can read the mark, it becomes “judge, jury, and prosecutor”. Anthropic told Business Insider it will ship a free detection API so anyone can check. On 14 August it said it is still working out the implementation details.

Former Microsoft executive Steven Sinofsky raised a different worry. The issue, he wrote, is “data retention and your right to private thoughts free of a digital trail”.

Simon Smith, who runs generative AI at the health agency Klick, asked whether a grammar check would now be flagged as AI-authored. Anthropic’s answer has two parts. The mark shows processing rather than authorship, and a grammar check will most likely carry no detectable mark at all.

The software trainer John Crickett asked a sharper question about code. If AI-generated code carries a mark, does that complicate a copyright claim in which the author must show human input?

That one now has an answer too. Code generally carries less marking than prose, because it has to be exact. The mark can attach to arbitrary choices such as comments, and by design it has a negligible effect on the code itself.

The case for doing it anyway

The objections are not the whole story. Developer Donn Felker pointed out that marking helps models avoid training on their own output, the problem he called a “snake eating itself”.

Aadit Sheth of The Narrative Company made the reader’s case. Audiences should be able to tell whether the words they are reading reflect a person’s thinking.

The Commission’s own reasoning is broader still. It wants people to calibrate their trust, and it names fraud, impersonation and consumer deception as the risks.

Anthropic frames this as compliance rather than a stance. It is adding marking “to comply with the EU AI Act, and other labs are taking similar steps”, the company said. It is applying the mark globally because it does not yet have a durable way to scope it by region, and says it will keep evaluating other approaches.

Some questions remain open. Ars Technica asked the company for a timeline on detection, for any testing on false positives and false negatives, and for how the marks square with the standard-editing exemption.

Anthropic sent a statement that did not address those points. Its 14 August post has since answered two of them, describing the failure modes and confirming the detection API is coming. It has still published nothing on false positives or false negatives.

That is the number that matters most here, because it decides how much weight a teacher or an editor should give the mark.

It will not stop anyone who is trying

The uncomfortable part is that the mark catches the honest and misses the deliberate. Paste watermarked text into a second model and ask for a rewrite, and the signal is likely gone. Anthropic agrees, saying a light edit probably will not remove the mark and a full rewrite will.

File-based provenance is weaker still, and it is a separate system. Files Claude produces carry a signed content credential in their metadata under the open C2PA standard, and nothing inside the file changes. A screenshot removes it, and so does any competent metadata editor.

The desk has seen this failure before. Meta’s own AI detector missed cropped images, which is roughly the amount of effort involved.

The Nature paper is candid about the rest. Generative watermarks can be stolen, spoofed and scrubbed, and paraphrasing weakens them. They also require the company running the model to apply them, which open-weights releases make close to unenforceable.

Anthropic is the only one doing text

This is the commercial risk, and it is real. OpenAI declined to comment specifically to TechRadar and pointed to a support page instead.

That page confirms OpenAI marks images and audio using SynthID and C2PA metadata. On text, it says expanding provenance to all modalities is a goal, as standards and tooling “continue to mature”.

Google is the sharper case. It invented the method Anthropic is using, published it in a peer-reviewed journal, and has been running it on Gemini since 2024. It has not shipped text marking for the European rules either.

Existing models have until 2 December to comply. Anthropic says marking for its own older models will roll out over the coming months. Until then, one lab is carrying the friction alone.

What the law does not touch at all

Article 50 requires disclosure that you are speaking to a machine. It says nothing about what happens to the conversation afterwards.

Stanford researchers reviewed the privacy policies of six major American AI developers. Every one used customer chat data to train by default, and some retained it indefinitely.

So the law now marks the output and ignores the input.

What would settle it

Three things, and all of them are checkable. The first is the detection API, and specifically its false-positive rate, because a mark on a student’s proofread essay carries the same weight as one on a fabricated news story.

The second is whether OpenAI or Google ship text marking before 2 December, or discover that the standards are still maturing.

The third is September, when the Commission opens two task forces for the roughly 190 organisations that signed its transparency code. That is the first venue where these approaches get compared side by side.

Labels already misfire without any of this. X put a “Made with AI” tag on the resignation post of the departing White House press secretary on Wednesday, and then the tag disappeared.