Claude adds invisible watermarks to text and files

How do you prove a block of polished writing came from an AI chatbot instead of a human?

Anthropic just answered that question on August 11, committing to machine-readable watermarks for text and files across its Claude ecosystem to comply with the European Union's AI Act.

Under Article 50(2) of the EU regulations, which went into effect on August 2, generative AI makers must make machine-made content identifiable. Anthropic is taking that obligation global.

Watermarking text without ruining the prose

Unlike traditional digital watermarking, this system does not stuff hidden Unicode characters into sentences, nor does it burn extra tokens or drive up API bills.

Instead, Anthropic borrowed a technique introduced in Google DeepMind's SynthID-Text paper. It works by tweaking how the language model rolls the dice during token generation.

When Claude predicts the next word in a phrase like "the weather was cold and...", words like "overcast" or "grey" are equally sensible. Normally, a random number generator picks between them. With the watermark active, that randomness is guided by a pseudo-random sequence tied to a private cryptographic key.

Anthropic compares it to playing Monopoly: instead of rolling ordinary dice, players take moves based on the digits of pi. To players, the moves still look totally random. But anyone holding the mathematical key can audit the sequence of moves and prove the game was using pi all along.

Anthropic says internal blind tests and human ratings showed zero drop in writing quality, tone, or creativity. To the human eye, the output looks identical to un-watermarked prose.

Global rollout, but detection stays behind closed doors

The marking rule applies to all new models launched on or after August 2, including across Claude Platform API, Claude Code, Claude Cowork, Claude Tag, and enterprise endpoints on AWS, Google Cloud, and Microsoft Foundry. Heavy hitters like Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 already support text watermarks, with older models catching up over time.

For visual files like PNGs and JPGs, Claude attaches signed metadata following the C2PA open standard, verifiable via a free web tool.

However, the text watermark detector is currently trapped in private preview. It is only accessible to regulators, law enforcement, researchers, media organizations, and enterprise compliance teams.

Anthropic is also quick to admit the system is fragile. Paraphrasing, heavy editing, or translating the text will degrade the watermark signal, as will very short sentences. Meanwhile, taking a screenshot or re-saving an image instantly strips C2PA credentials.

Why it matters

For readers and online communities, built-in tracking is a welcome relief. On platforms like AO3, fanfiction communities have spent months jury-rigging crude scrapers to catch AI text, while Substack recently partnered with detection tools after its CEO sounded the alarm on "Claudefishing." But if you rely on Claude for everyday proofreading, there is a catch: running your own drafted work through the model could slap an AI tag onto your writing.

For the AI landscape, this shows how aggressively EU regulations are reshaping product design. Tech giants from Google and Meta to OpenAI and Black Forest Labs have signed the same transparency code. Marking content is no longer a niche safety experiment; it is the table stakes for doing enterprise business in Europe.

We think Anthropic was smart to embed the watermark at the token sampling layer, keeping prose natural while resisting simple copy-paste scrubbing.

That said, keeping the detection API locked away leaves the everyday web in the dark. As long as users cannot verify text themselves and bad actors can bypass the tag with a few quick edits, this looks more like regulatory armor for tech giants than a bulletproof shield against online spam.