Anthropic has offered a clearer look at how Claude will mark AI-generated text without adding a visible label. The company says the approach relies on subtle patterns inside wording choices, creating a signal that readers do not see but that can be detected by someone with the right key.
The move is tied to Europe’s AI transparency rules. It also shows how major AI developers are preparing for a world where generated content is expected to carry machine-readable evidence of its origin.
What Anthropic says it is adding to Claude
On Friday, Anthropic announced that Claude’s text marking system is “a version of the SynthID-Text approach.” SynthID-Text is an open-source watermarking technology developed by Google DeepMind.
The core idea is not to stamp a visible warning across the text. Instead, the system creates detectable patterns through the probabilities behind word choices. Those patterns sit inside the generated text itself.
Anthropic is pairing this text watermarking feature with C2PA support for Claude-processed images. Together, those tools are being introduced so the company can meet obligations under the European Union’s AI Act.
According to the source article, the European Union’s AI Act requires synthetic audio, image, video, and text to include machine-readable marks. Those marks are intended to make it possible to detect content that was artificially generated or manipulated.
How invisible text watermarking works
Anthropic’s explanation centers on the way a language model picks among likely next words. In many sentences, there are several reasonable options that can preserve the same general meaning.
The company uses a simple example: after the words “The weather today was cold and…”, a word like “sugary” would be very unlikely. Words such as “overcast” or “grey” would be much more plausible. For a reader, either of those likely choices can communicate a similar idea.
In ordinary generation, a random number can help settle which acceptable word is chosen. In the watermarking process, Anthropic says the randomness still exists, but its source changes.
Instead of using an arbitrary random number generator, the system uses a key and several preceding words to guide the choice. Repeated across a longer piece of generated text, those small decisions can form a pattern.
The result is meant to be invisible in normal reading. A person reading the paragraph should not notice the watermark, but someone with the key that encodes the pattern can detect it.
Why the watermark is meant to stay out of the reader’s way
Anthropic says the system is built around low-stakes wording choices. These are moments where one likely word can be selected instead of another without meaningfully changing what the output says.
That distinction matters because text does not have pixels or audio signals where a hidden mark can be placed separately from the words. For generated writing, the mark has to be created through the generation process itself.
Anthropic says the text watermarks will not make Claude more expensive for users. The company also says they will not “have any practical impact on the quality or content of Claude’s outputs.”
In plain terms, the goal is to make the generated text machine-detectable while keeping the user-facing result the same. The reader sees ordinary prose. The detection system looks for a statistical pattern in the choices behind that prose.
What this means for the wider AI market
Claude is not the only AI system affected by Europe’s AI transparency requirements. Anthropic notes that other major AI developers are also covered by the rules.
Google’s Gemini chatbot has supported the SynthID-Text solution since 2024. That makes Claude’s move part of a broader shift toward watermarking and provenance tools for AI-generated content.
OpenAI has not detailed any text watermarking plans for ChatGPT in its AI Act compliance roadmap, according to the source article. Even so, it will also be subject to the law’s requirements.
The practical stakes are straightforward:
- For AI developers: generated text needs a way to carry machine-readable signs that it came from an AI system.
- For users: Anthropic says Claude’s watermarking should not change the cost, quality, or content of outputs.
- For detection: the mark depends on patterns that are invisible to readers but visible to someone with the right key.
The broader question is how consistently these systems will be adopted across AI tools. The source article makes clear that the same legal pressure reaches beyond Claude, while also showing that companies may disclose their text watermarking plans at different speeds.
The shift from visible labels to machine-readable proof
Claude’s planned watermarking points to a more technical form of AI transparency. Instead of relying only on a visible notice, the system is designed to place evidence inside the generated material.
That evidence is not meant to be read by people directly. It is meant to support detection, especially where synthetic or manipulated content needs to be identified by machines.
For now, Anthropic’s explanation gives a basic model of how this can work: choose among similar words in a patterned way, repeat that process across the output, and allow detection through a key. If it works as described, Claude’s text can carry a hidden signal while still reading like normal generated text.