What Is AI Text Watermarking?

AI text watermarking embeds an invisible, machine-readable signal into text produced by an AI model, so the text can later be identified as machine-generated.

You cannot see it. It is meant to survive copying and pasting. It became mainstream in August 2026 when Anthropic began watermarking Claude’s output.

What a text watermark actually is

Not a visible label, and not metadata you can strip by pasting into a plain text editor. The signal is carried in the text itself. Anthropic describes it as imperceptible and says it does not affect quality or readability.

For files rather than raw text, providers use C2PA, an open standard for recording content provenance.

Why AI watermarking started

Regulation, primarily. The EU AI Act’s Article 50(2) Code of Practice on Transparency took effect on 2 August 2026, requiring providers of generative AI to mark their output.

Anthropic signed that code and applied watermarking globally, not only for European users. Read more about the EU rules.

The wider driver is what the press calls AI slop: the volume of unlabelled machine-written material now circulating, and the pressure on platforms to tell the difference.

How watermarking differs from AI detection

These get confused constantly. They are not the same thing.

• A detector analyses writing patterns and outputs a probability. It can be wrong in both directions.

• A watermark reads an embedded signal. It is present or absent, with no percentage.

• A detector works on any text. A watermark only exists in text from participating models.

• Absence of a watermark proves nothing at all.

A detector guesses whether writing looks machine-generated. A watermark records that text passed through a specific model. That is why watermarking does not replace AI detection.

What a watermark can and cannot prove

It can indicate: this text was processed by a particular model.

It cannot establish: who wrote the underlying ideas, whether any policy was broken, whether the use was legitimate, or whether a person edited and fact-checked the output.

Plenty of legitimate uses put text through a model: translation, accessibility tools, grammar correction, or reformatting something you wrote yourself.

What is still unresolved

• How much editing degrades or removes a watermark is not publicly documented

• It is unclear who will be able to check for watermarks, and through what tools

• Other major providers have not announced equivalent text watermarking

• How institutions will treat a watermark in policy terms is undecided

Last checked 12 August 2026.

What Lunchbreak does

Two tools, both aimed at the moment before you hand something in.

The AI checker lets you paste your writing and see how major AI detectors read it, side by side. Because detectors are probabilistic, they often disagree. Seeing several results together tells you more than any single score.

The humanizer rewrites for tone, rhythm, and flow so writing sounds like a person rather than a template. You see the original and the rewrite together and choose what to keep.

Neither tool reads watermarks. Watermark checking requires a tool built for a provider’s specific signal, and no public one exists yet. What Lunchbreak works on is the other half of the problem: how your writing reads, and whether it sounds like you.

That is worth doing regardless of watermarking. Detector false positives affect human writing too.

Check my writing free