Anthropic announces that future Claude models will generate text with an invisible watermark. The goal is to calculate how likely it is that Claude participated in creating a piece of content, without altering its quality, adding hidden characters, or increasing its cost.
The measure responds to the European Union AI Act, which requires providers to identify content generated by artificial intelligence. Anthropic says it will apply the system globally because it still does not have a durable way to limit it by region.
How Claude’s Watermark Works
Claude generates text by choosing one word or fragment at a time. In many cases, there are several equally valid options. For example, after writing “The weather today was cold and...,” the model might choose “cloudy” or “gray” without changing the meaning of the sentence very much.
The watermark takes advantage of precisely these low-risk decisions. Instead of using a completely arbitrary source of randomness, the model uses a key and some previous words to guide those choices. The result is a statistical pattern that people cannot distinguish, but that can be analyzed with the right key.
Nothing is added to the text. There are no invisible characters, hidden tags, or visible signals for the reader.
The technique is based on a version of SynthID-Text, a method introduced by Google DeepMind in a paper published in Nature in 2024. Its principle is simple: preserve the options the model would already consider reasonable and change only how it selects among them.
Does the Watermark Change the Response?
According to Anthropic’s internal tests, the watermark does not affect the quality, creativity, or readability of responses. A reader should not be able to distinguish between marked and unmarked text.
The company also cites research by Google DeepMind with Gemini. In those tests, marked and unmarked responses received similar ratings, with no statistically significant differences in users’ positive or negative opinions.
One way to understand it is to imagine a board game. If players move using dice, there is a sequence of seemingly random results. But if you replace the dice with a sequence of pi digits, the moves can still appear random. Someone who knows the sequence could review the game and calculate whether pi was probably used.
The same thing happens with Claude’s text: the reading experience does not change, but an authorized system can review the pattern of choices and estimate whether Claude participated in the generation process.
What the Watermark Can Detect
The system does not absolutely prove that Claude wrote a text. It only calculates the probability that Claude was involved at some point. It also cannot confirm that the content is human-created or identify whether it was created by another AI model.
Accuracy depends, among other things, on the length of the text. A longer passage contains more decisions between possible words and provides more information for analysis. A short sample, on the other hand, may not include enough choices for detection to be reliable.
The watermark is also weaker in texts where there is only one correct answer. In a sentence like “Isaac Newton’s most famous work was Principia...,” the next word must be “Mathematica.” Changing that choice would make no sense because it would reduce accuracy.
Editing, Translation, and Text Correction
When Claude corrects the grammar of a human-written text, most of the words still belong to the original writer. If it makes only a few changes, there may not be enough material to detect the model’s participation.
The situation is different with a complete translation. Because Claude chooses practically every word in the result, the content can carry a more detectable watermark.
Light editing probably will not completely eliminate the signal. However, a total rewrite, in which every word is replaced, can neutralize it. At that point, another reasonable question arises: is it still accurate to describe that text as AI-generated?
What Happens with Code?
Code tends to contain fewer watermarks because many of its decisions must be exact. If the model writes “2 + 2 =”, the correct answer is “4.” There is no equivalent alternative the system can use without risking an error.
The watermark may appear in areas with more freedom, such as code comments or certain descriptive names. Even so, Anthropic says its effect on functional code will be negligible.
What It Means for Users
The technology does not require additional tokens and will have a negligible impact on response speed. Anthropic also says it will not increase the model’s usage or service price.
The watermark does not contain information about the person who used Claude, their organization, or their conversations. It also cannot trace a text back to a specific user.
Anthropic plans to offer a detection API soon to check whether a text was probably generated or processed by Claude. The company is still defining the details of that tool.
For images and other compatible files, such as .png, .jpg, or .svg, Claude will use C2PA content credentials. This is an open standard that adds a cryptographically signed note to the file’s metadata to indicate that Claude participated in its creation or processing.
This does not modify the image or insert a hidden signal. The credential also does not include data that identifies the user. Tools compatible with C2PA will be able to read it.
Why Anthropic Is Implementing This Measure
Anthropic explains that the decision aims to comply with the European Union AI Act. The company and other major providers signed the Code of Practice on Transparency of AI-Generated Content in July 2026, along with nearly 190 organizations.
The watermark will not be exclusive to Claude. Other model developers are also preparing their own identification systems, although each one may use a different key and method.
Claude models released before August 2, 2026, have a transition period. Anthropic is working to add the watermark to them over the coming months.
A Useful Signal, but Not Definitive Proof
The proposal attempts to solve a specific problem: determining whether a model participated in creating content without harming the person reading it. But it does not turn detection into an automatic judgment about authorship, ownership, or legal responsibility.
A watermark can only indicate that Claude probably took part. It does not determine who the author is, who owns the result, or who should be responsible for its use. At a time when AI is already part of writing, translation, programming, and editing, that distinction will become increasingly important.