Anthropic has officially explained how Claude's text watermark works, adopting Google DeepMind's SynthID-Text technology to comply with EU requirements for machine-readable marking of AI-generated content. The watermark will be applied to all Claude users globally, and while the company claims negligible impact on output quality, the article raises questions about long-term effects.
SynthePulse Insight · AI deep reading
Version 1 · 2 sources
Anthropic has announced the addition of invisible watermarks to Claude-generated text, using Google DeepMind's SynthID-Text technology, to comply with the EU AI Act. This article delves into the principles, impact, limitations, and future detection tools of watermarking.
Anthropic announced Claude's content marking scheme on August 11, 2026, via the X platform, and then explained the workings of text watermarking in detail through an official blog post on August 15. The core principle of the watermark leverages the random choices made by large language models during text generation: when the model chooses among multiple equally suitable candidate words, the choice, originally determined by a random number, is instead determined by a secret key and preceding words, leaving a detectable pattern in the text.
This pattern is invisible to readers, but those holding the key can detect it. Anthropic emphasizes that the watermark does not force the model to choose words it would not otherwise consider, such as suddenly using obscure words like 'nubilous'. The watermark only changes the source of randomness, not the meaning, quality, or readability of the text.
Anthropic chose to adopt a version of Google DeepMind's SynthID-Text method, published in Nature in 2024. This line of methods traces back to a proposal by Scott Aaronson in 2022. The primary purpose of implementing watermarking is to comply with the EU Artificial Intelligence Act, which requires AI providers to mark AI-generated content in a machine-readable way.
In July 2026, Anthropic, along with other major AI providers and a total of approximately 190 signatories, signed the EU's AI-Generated Content Transparency Code of Conduct. Due to the current lack of reliable regional differentiation schemes, the watermark will be applied globally, not just in the EU.
Anthropic claims that watermarking has no impact on Claude's output quality, and internal tests found no effect on content, creativity, or readability. In Google DeepMind's SynthID-Text paper, comparisons using user upvote/downvote scores and human raters found no statistically significant differences.
Watermarking does not generate extra tokens, so it does not increase service costs or user prices, and the impact on model speed is negligible. However, the blog uses many qualifiers such as 'largely the same', 'no statistically significant difference', and 'negligible impact', implying that it is not absolutely without effect. APPSO's report also points out that 'almost no impact' is not the same as 'no impact at all', and in long-term use, generation quality might be sacrificed in certain scenarios.
Watermark detection is ineffective for short texts because there are few word choices and insufficient information. In more factual texts, watermarks are sparser because there are fewer freely selectable words, and forcing changes could affect factual accuracy. For example, in mathematical calculations or code, precise output positions will not be watermarked.
If Claude only proofreads or edits human text, the watermark may attach little or none, because most words come from the original author. Fully rewriting every word can remove the watermark, but light editing may not completely remove it. The watermark can only determine whether Claude is likely involved, not whether it wrote the text or heavily edited it.
The watermark does not carry any identity information and cannot be traced back to specific individuals, organizations, or chat records. It does not change ownership or legal responsibility; it is only used to determine whether content was generated or processed by Claude.
Anthropic is about to launch a watermark detection API, but implementation details are still being finalized. For files such as images, Claude will attach C2PA content credentials in the metadata, which is an open industry standard and not embedded in the content itself. Watermarks for older Claude models will be added gradually in the coming months to comply with EU transition period requirements.
Watermarking differs from AI detection software like Pangram. The latter does not have the key and detects by analyzing text features such as common AI sentence patterns (e.g., 'It's not X, but Y') and word frequencies (e.g., 'quietly'). Watermarking, on the other hand, is deterministic detection based on a key, with a completely different principle.
This report is based on Anthropic's official blog and X platform announcements, as well as APPSO's retelling. The official blog is a primary source, but some statements (such as 'no statistically significant difference') come from Anthropic's own claims and have not been independently verified. APPSO's report is a secondary source, and its commentary is media opinion. The actual impact and detection effectiveness of the watermark still require independent verification in the future.
Claude text watermarking is a technology introduced by Anthropic to comply with the EU AI Act, using the SynthID-Text method. It has minimal impact on output quality but has detection limitations. Users need to be aware of the impact on editing, translation, and other scenarios, as well as the actual effectiveness of future detection tools.
Primary report
Primary source
Also covered by 2 sources