Claude is going to leave an invisible mark in its text: what it really means

I'm reviewing what Anthropic published on August 14 about Claude's watermark because the headline making the rounds is quite a bit more aggressive: “Claude will mark everything it writes.”

The true part is important. Compatible Claude models will embed an invisible signal in text that makes it possible to estimate whether Claude took part in generating it. The part being told wrong is the mechanism. There are no secret characters hidden between words, no user identifier is added and there is no label saying which conversation the text came from. Anthropic

The mark lives in the way the model chooses certain words.

The mark isn't hidden characters

A language model builds a response word by word. At certain moments it has several reasonable options that keep practically the same meaning.

Anthropic gives the example of choosing between “overcast” and “grey” to continue a sentence about the weather. When both options are valid, the model normally introduces some randomness to decide.

With the watermark, that randomness is controlled by a key. Over a sufficiently long text, the choices leave a statistical pattern. A detector that knows that key can check the pattern and calculate how likely it is that Claude took part. Anthropic

Anthropic confirmed that its implementation is a version of the SynthID-Text approach published by Google DeepMind. Anthropic

The signal is in the model's statistical decisions, not in characters added to the document.

That explains why copying and pasting the text can preserve the mark. The pattern is formed by the words chosen. A light edit can weaken it, but not necessarily remove it. Anthropic

What it can actually prove

Here is the part that matters more to me than the headline. A positive detection is not the same as proving authorship.

Anthropic also clarifies that the mark contains no information about the person, the organization or the chat where it was generated. Anthropic

That changes the interpretation quite a bit. A marked document could have been originally written by a person and then passed through Claude to be translated or edited. The detector can indicate model involvement; it doesn't reconstruct the document's full history.

In code the signal is more limited

For those of us who use Claude Code, this part deserves attention.

The watermark works best when the model has several equally valid words. In code the opposite is often true. A reserved word, a property name or the result of an operation may need an exact form.

Anthropic says that in those cases the mark doesn't force an alternative. That's why code usually has less watermarking than prose. Comments and parts with more freedom do offer more room to introduce the signal. Anthropic

That seems like a reasonable technical decision to me. I'd rather a provenance mechanism have less signal than alter code to make it easier to detect.

Where it will apply

Anthropic is applying the system at the model level, not as a feature exclusive to Claude's interface.

For compatible models, the marking covers:

• Claude on web and apps.
• Claude Platform and the API.
• Claude Code, Cowork and other Claude surfaces.
• Access through AWS, Google Cloud and Microsoft Foundry when the model supports the mechanism. Claude Help Center

The company states that models released from August 2, 2026 onward support marking from launch. Earlier models are in the process of receiving support, so today it's not wise to claim that absolutely every historical Claude response already carries a watermark. Claude Help Center

Files work differently.

When Claude produces or processes compatible types such as PNG, JPG or SVG, it can attach C2PA-based provenance metadata. It is a signed credential that records Claude's involvement. That information lives in the file's metadata and can be lost when converting formats, re-saving an image or passing it through platforms that strip that data. Anthropic

Text and files use different mechanisms.

Why Anthropic is doing it now

The immediate origin is regulatory.

Article 50 of the European Union's AI Act began to apply on August 2, 2026 and sets transparency obligations for AI-generated or manipulated content. The European Commission also published a Code of Practice that serves as a way to demonstrate compliance. European Digital Strategy

Anthropic signed that code along with other providers and decided to deploy the marking globally because, according to the company, it doesn't yet have a durable way to limit it to a single region. Anthropic

This means a European requirement ends up affecting users outside Europe. If the model supports the watermark, the signal can appear even if you're working from Mexico.

The discussion about quality does make sense

There's an interesting technical controversy here.

Anthropic maintains that the watermark has no practical impact on quality, creativity or readability. Its explanation is that the system acts only when several suitable options exist and changes the source of randomness used to select among them. Anthropic

The technique is related to SynthID-Text, which Google DeepMind already tested with Gemini traffic without finding statistically significant differences in user ratings, according to the explanation published by Anthropic. Anthropic

The criticism is conceptual: if you modify the way a model chooses words, there's a chance some selection will be stylistically less precise than another. John Gruber was especially harsh on that idea and argued that interfering with word choice affects the act of writing. Other specialists consulted by The Guardian believe the effect will probably be imperceptible because models already rely on randomness during generation. The Guardian

For now, the evidence presented by Anthropic points to a minimal impact. It also remains to be seen how the system behaves outside internal tests and with specialized, creative or heavily edited texts.

My takeaway

There are four things I would change in how I work with Claude:

• Stop treating detectors as a verdict on authorship.
• Document when an AI took part in important professional deliverables.
• Separate the concept of statistical watermark from C2PA metadata.
• Review internal policies if a client has restrictions on AI-assisted content.

The mark can be useful for provenance, but reading it needs context. “Detected” doesn't mean “written entirely by Claude,” and “not detected” doesn't mean “written by a person” either.

What I would do on a development team

If Claude is part of the daily workflow, I wouldn't try to build a process around “erasing the mark.” That turns a transparency matter into a race against detectors that also have technical limitations.

I'd do something simpler: define which uses of AI are allowed, which deliverables require human review and when it must be recorded that a generative tool was involved.

For code, I'd keep exactly what should already exist: review, tests and human responsibility for what reaches production. For documents, proposals or public content, I'd add a clear rule about provenance when the legal or contractual context requires it.

Watermarking technology will keep changing. The useful part today is understanding what signal it offers and what conclusion we can't draw from it.