# Claude's Watermark: Are We Measuring the Wrong Thing?

> Source: <https://dev.to/eleonorarocchi/claudes-watermark-are-we-measuring-the-wrong-thing-1ahj>
> Published: 2026-08-17 05:53:00+00:00

Since August 2, 2026, the [EU AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX%3A02024R1689-20260727) has required certain AI-generated content to be detectable through machine-readable markings.

Anthropic has chosen a technique for Claude derived from **SynthID-Text**, developed by Google DeepMind. The mechanism does not insert invisible characters: it slightly modifies the token-sampling process, leaving behind a detectable statistical signature. Anthropic explains it here: [How Claude's text watermark works](https://www.anthropic.com/news/claude-text-watermark).

Detection is probabilistic and works less reliably on short, highly factual texts, code, or lightly edited content. Google documents these limitations in the [SynthID documentation](https://ai.google.dev/responsible/docs/safeguards/synthid).

But to me, the most interesting part is not the technical one.

A watermark can tell us that Claude probably played a role in the **wording** of a text. It cannot tell us who came up with the idea.

And that is a huge distinction.

I do not see a problem with using AI to edit, summarize, translate, or make content clearer. After all, we have always had editors, proofreaders, and ghostwriters.

The problem arises when we delegate not the wording, but the **thinking**.

If someone has a thesis, chooses the arguments, verifies the sources, and uses Claude to express them better, I still consider that thinking their own.

If, on the other hand, they ask AI to produce the thesis, arguments, and conclusions without exercising any judgment, then there is a real problem.

But not because AI wrote the words.

Because it replaced the reasoning.

That is why I find the generic label **"AI-generated"** increasingly unhelpful: it puts a grammatical correction and a text entirely delegated to a model in the same category.

More importantly, it risks making us focus on the wrong thing.

AI is driving down the cost of content production. Writing well, summarizing, translating, rephrasing: all of this is becoming abundant.

What remains scarce is **having something to say**.

That is why, in a world of infinite content, the important question may no longer be:

"Did you use AI to write this?"

But:

"Is this thought really yours?"

A watermark can help us understand where the words came from.

Not where the ideas came from.

And perhaps that is the distinction that really matters.