Anthropic is adding invisible watermarks to text produced by its Claude AI models as part of efforts to meet European Union rules about AI transparency. These marks are not visible to readers but help identify when content was created with the help of an AI system.
The change applies to models launched in the EU after August 2, 2026, and will also be rolled out to older versions, according to ibtimes.com. The European Commission requires that AI-generated or manipulated content can be detected using machine-readable markings.
Watermarking AI-generated text is a method designed to help identify whether written content was produced by a generative system. Unlike a visible watermark on an image, a text watermark is usually hidden within patterns of word choice, sentence structure, or token selection. An AI system can be programmed to favor certain statistically detectable word sequences without noticeably changing the meaning of the text. Specialized detection software can then analyze a passage and estimate whether those patterns are present. Researchers have explored watermarking as a way to identify automated content, reduce misuse, and make AI-generated material easier to trace.
Not every piece of AI-assisted text needs to show a clear warning to readers, but systems must allow detection. The technology used by Anthropic is based on research from the 2024 SynthID paper and shares similarities with Google’s own system, according to a report from Gizmodo.
Watermarking works by using a secret key that influences token choices in the text, which are statistically preferred but not noticeable to humans. These low-stakes choices are context-dependent and can be detected only through a specialized system or API.
In longer texts, enough of these tokens form a detectable watermark, but shorter content may not show the mark clearly. The system does not slow down text generation or increase token costs significantly.
Even if someone uses Claude to edit human-written text, the watermark may still appear depending on how much the AI changes it. Some users have canceled their subscriptions in response to this change, expressing frustration and disappointment.
Text watermarking has important limitations, however. Editing, translating, summarizing, or paraphrasing a passage can weaken or completely remove the detectable pattern, and short passages may not contain enough information for reliable identification. A watermark can also produce false positives if naturally written text happens to contain similar statistical patterns. Because of these weaknesses, watermarking is generally considered more useful as one piece of evidence rather than definitive proof that a person used AI to create a document. Researchers are continuing to study stronger methods that can survive editing while avoiding changes that make AI-generated writing sound unnatural.
One well-known AI influencer said he canceled because he felt the watermarking idea was ridiculous. Users have posted negative reactions on social media platforms like X, calling the move dishonest or offensive.
However, not all users are leaving Anthropic, and some cancelations may be due to unrelated issues like CEO Dario Amodei’s wife’s business history. The system is designed to prevent people from pretending they wrote AI-generated content themselves.
Older Claude models that were already on the market before August 2, 2026, have until December 2, 2026, to comply with the new marking rules. This effort reflects a broader shift in how companies approach AI content transparency and regulation compliance.
The invisible nature of the watermarking system means it won’t change how text appears or reads to users, NewsBytes reported. Despite backlash, Anthropic claims that the number of cancelations has not increased since watermarking was announced.
The company says its approach is in line with industry standards and necessary to meet EU guidelines. Users who rely on AI for editing or assistance may find their work flagged even if they believe it’s original.
Some users are concerned that the watermarking will be used by employers or schools to detect AI use in submitted work. The system is not perfect and may fail in cases where text is heavily rewritten or combined with other material.
Overall, the move shows how AI companies are adapting to new legal requirements while balancing user expectations.
IMAGE: fill / Pixabay