A watermark can't tell you who's responsible

On August 11, Anthropic announced that text generated by Claude will now carry an invisible watermark. It doesn't affect reading, can't be seen with the naked eye, travels with copy-paste, and may survive light editing. Images work differently: PNG, JPG, and SVG files get digitally signed provenance metadata, which drops off with format conversion or screenshots. It applies everywhere — regardless of region, and regardless of whether the output came from the web app, the API, or Claude Code.

Once the news broke, the backlash was loud. Paying users felt treated like suspects, novelists worried their submissions would get flagged, and freelancers worried clients would start demanding deliverables with "no traces of AI."

There's only one reason to be angry

A photo from a camera already carries the body model, lens, aperture, and ISO in its file. No photographer feels their work is tainted by this, and no one argues the photo's author is Canon. Word has spell check; no one worries about being caught using it. Tools leaving traces on finished work is entirely ordinary.

So why does a text watermark set people off?

I can think of only one reason: fear of being detected. And fearing detection presupposes that, deep down, you feel the work isn't entirely yours. If you genuinely believe AI is just a tool, that the content is yours, and that the responsibility is yours to carry, a watermark is no more worth worrying about than EXIF data.

Many people say AI is a tool. They don't actually believe it.

Why they did it

The immediate reason is regulation. Article 50 of the EU AI Act imposes transparency obligations on AI-generated content, and it took effect on August 2. Anthropic signed the corresponding code of practice, which means any new model launched in the EU must support marking from day one.

The regulation covers only the EU; Anthropic chose to apply it globally. Technically, the watermark is added at the model level — splitting by region would mean maintaining two models with different behavior, which costs more than doing it everywhere. Commercially, the first companies to sign publicly get remembered, and Anthropic's course in recent years has been to cooperate proactively with regulators and trade compliance for trust. This move fits that posture.

There's also a layer about who gets to write the rules. C2PA, the file provenance standard, was pushed forward largely by Adobe — the early movers decided what verification looks like and who is qualified to sign. Text watermarking has no accepted approach yet, so whoever moves first has a chance to turn their method into the de facto standard.

The detection tool still hasn't been released, and the technical details haven't been published. For now, the only party that can verify this watermark is Anthropic itself. They say it will be offered to third parties later, but until then, this is a key held by exactly one person.

The people who built the watermark say it can't identify authors

There's a remarkably honest passage in Anthropic's announcement: detecting the mark only means the content "may have" passed through Claude — not that Claude is the original author. People routinely use it to proofread, translate, summarize, and convert files, so the underlying ideas, data, and sentences may come from somewhere else entirely. Conversely, the absence of the mark doesn't mean the text wasn't written by AI.

The company that built this thing says, in its own words, that it cannot establish authorship. They haven't even published the detection tool, the false positive rate, or the minimum text length for reliable detection.

What happens next

Anthropic said all of this clearly. Nobody downstream will read it.

Schools, publishers, contest organizers, and HR departments will inevitably start using this as evidence. A statistical signal will be treated as a verdict, because it saves work. A reviewer no longer has to read the manuscript and judge whether it's any good — run a detector, and the decision makes itself.

The step being saved is responsibility itself.

When a piece of writing goes wrong, the people to hold accountable are whoever decided to send it out, whoever signed their name to it, whoever took the money. None of that has anything to do with whether a tool left traces. Responsibility has never been something you detect — it's something you carry. Replacing judgment with a detection result means outsourcing the question of "who's responsible" to a machine — a machine whose maker has already said it can't answer that question.

For freelancers

What will change day to day is the acceptance criteria.

If clients start demanding deliverables with "no traces of AI," then what's being judged shifts from whether the content is good to whether its origin is pure. These two things are unrelated. A bad idea doesn't become good because it's handmade, and good copy doesn't stop working because a machine was involved. Spending effort scrubbing away traces is spending effort in the wrong place.

The work I deliver carries my name. I'm the one the client yells at, the one who does the revisions, the one who eats the loss. None of that changes because the file carries an extra invisible layer of marks.

One last thing

Text watermarking works by quietly nudging the probability distribution the model uses to choose words, giving the whole text a statistical tilt. It's a deliberately manufactured difference — but one that carries no meaning. It doesn't change what any word means or what any sentence says. Its only job is to point to a source.

I've always believed meaning comes from difference. The watermark is an exception — a difference drained of semantics, used purely as a signature.

What they solved is provenance, not responsibility. Conflating the two is where the trouble will come from over the next few years.

I'm not against watermarks. I'm against using them as an excuse not to judge.

(Part of this piece was written with Claude, so it probably carries a watermark. The name on it is still mine.)