Claude’s Invisible Watermark Will Follow AI Text Everywhere

Anthropic is adding an invisible watermark to text produced by its Claude models, creating a way to identify that output without changing how it looks to readers. Claude models launched on or after August 2, 2026, will include an “imperceptible watermark” in their text output.

The move follows European Union transparency regulations for artificial intelligence. Anthropic plans to add the same watermarking capability to older Claude models over the coming months, with watermarking applied to all Claude output after August 2 and added to older models in stages.

The watermark does not appear as a symbol, image, or strange character. It also does not add hidden characters to the text. Instead, the mark is a pattern created by the model’s word choices, which means it travels with text when someone copies and pastes it.

How Claude’s Text Watermark Works

Claude chooses words based on probabilities, but Anthropic’s system adds a key to those choices. The key decides which word Claude selects instead of leaving the decision to an arbitrary random number generator. One example of a key is the digits of pi.

Those choices may look ordinary when viewed one at a time. Across a full piece of generated text, however, they create a pattern. Jess Weatherbed described the system this way: “Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it.”

The pattern can only be decoded by someone who has the key. Anthropic will release an application programming interface, or API, with keys for decoding Claude’s watermark. That gives the company a way to support detection without making the watermark visible inside the text itself.

Claude’s system uses a version of SynthID-Text, the open-source watermarking technology developed by Google DeepMind. Weatherbed described Claude’s text marking system as “a version of the SynthID-Text approach” and said it creates detectable patterns through wording probabilities.

What the Watermark Can and Cannot Show

The watermark can identify text that contains Claude’s pattern, but it cannot determine whether Claude wrote every word from the start or only edited text created somewhere else. That distinction matters because the method tracks the pattern in Claude’s word choices, not the full history of a document.

Lightly edited Claude-generated text probably will not lose its watermark. Copying and pasting the text also leaves the watermark attached, since the pattern exists in the wording rather than in a separate file, image, or visible marker.

Anthropic says the system will not affect the quality or content of Claude’s outputs. It also says watermarking will not change the speed of Claude’s output, require extra tokens, or make generations more expensive.

“Watermarking doesn’t affect the quality of Claude’s output or slow it down,” Anthropic said, adding that the method will not require extra tokens or make generations more expensive.

Anthropic’s Watermarking Rollout

Anthropic revealed how it was watermarking Claude text on August 15, 2026. The explanation followed an August 14 article titled “How Claude’s Text Watermark Works,” an August 17 explanation by Jess Weatherbed, and coverage published on August 18, August 20, and August 20 by Business Insider, CNN, and The Verge.

The rollout began with Claude models launched on or after August 2, 2026. Anthropic will then extend watermarking to older Claude models over the coming months, while providing decoding keys through its API.

For people handling AI-generated writing, the key detail is easy to miss: Claude’s watermark is not something a reader can spot by looking at the page. It is a coded pattern inside the model’s word choices, designed for detection by someone with the right key.

Stay connected via Google News

Follow us for the latest AI News