Hidden Marks: Anthropic Watermarks Claude's Text
If you use artificial intelligence to draft an email, a report, or a social media post, the text you produce may now carry a hidden marker. Anthropic, the company behind the Claude AI models, has started embedding an invisible, machine-readable watermark into new Claude text output. As of August 2, 2026, the watermark is part of all new Claude models, according to Proto Thema, which described Anthropic as the first of the major providers to implement such a system for its AI text.
The watermark is not a visible badge or a label reading "Written by AI." Instead, it is a statistical pattern in how the model selects certain words, detectable only by a specialized tool. To the reader, the text appears normal. Anthropic has chosen to roll out the watermark worldwide, not just for users in Europe, a fact reported by both Proto Thema and Kiplinger.
The Regulatory Driver: Europe's AI Act
The initiative stems from the European Union's Artificial Intelligence Act. Kiplinger reports that the EU's strict rules, which include a broad set of requirements for AI deemed high-risk, went into effect this month. Proto Thema cites Article 50 of the AI Act, which applies from August 2, 2026, requiring providers of AI systems that generate synthetic text, images, audio, or video to ensure content can be identified as artificially generated through machine-readable labeling.
The aim, as Proto Thema explains, is to reduce risks of misinformation, manipulation, fraud, impersonation, and consumer deception. The European Commission has said Europe wants to make it harder for content that appears human-made to circulate without a way to determine it was created by a machine.
The EU AI Act also prohibits certain uses of AI, including deploying subliminal, manipulative, or deceptive techniques that distort behavior and impair informed decision-making, as well as inferring emotions in workplaces or educational institutions, except for medical or safety reasons, according to Kiplinger. The most severe penalties for violations are fines of up to 7% of global revenue, Kiplinger notes.
How the Watermark Works
Anthropic's watermarking method, as detailed by Kiplinger, was developed and already used by Google. The process involves how the AI model chooses specific words and word fragments. Anthropic holds a key that involves two lists of words, and the generated text must include enough words from one list to be statistically significant for the watermark to be detected.
The watermark has limitations, Kiplinger reports: it does not work well on shorter passages and only reveals the "likelihood" that text was AI-generated, rather than providing certainty. Proto Thema adds that the watermark can survive copying, pasting, and light editing, but extensive rewriting, paraphrasing, or processing by another AI model can weaken or eliminate its statistical trace.
Because the watermark is applied at the model level, it is present in all Claude products and surfaces, according to Kiplinger. Anthropic states that readers will not notice the watermark and that it does not change the meaning, quality, or readability of responses, a claim carried by both outlets. Anthropic also maintains that the watermark does not increase costs or degrade quality, Proto Thema notes, though it acknowledges a discussion about a slight impact on quality.
Anthropic has not publicly released the detection tool for its watermark, Proto Thema reports. Meanwhile, tools are already appearing on the market that promise to make AI-generated text more human-like or to bypass detection systems, the outlet adds.
Industry and Political Reactions
Anthropic has said it is implementing watermarking to comply with the EU AI Act, and that other major model developers have signed the same Code of Practice and will also implement watermarking. Approximately 190 organizations, including Anthropic, Google, Meta, Microsoft, Mistral, and OpenAI, have signed the European Code of Practice on transparency of AI-generated content, according to Proto Thema.
Google, which developed its own SynthID watermarking system, is working on interoperable watermarking tools with other companies, Proto Thema reports.
The move is likely to stir concerns about Anthropic's power over users' text output and implications for intellectual property, Kiplinger suggests. It also expects pushback from the Trump administration, citing President Trump's statement last month that the administration will conduct a formal review to retaliate against the EU's "discriminatory" digital practices. Kiplinger predicts that this backlash will focus more attention on Google's use of text watermarks for its AI model Gemini.
A Regulatory Shift, Not a Feature
Europe is turning detectability of AI-generated content into a regulatory requirement rather than an optional feature, Proto Thema observes. The broader context, as the outlet describes, is an internet flooded with "AI slop"—vast amounts of cheap, hastily produced, low-quality content—making it increasingly difficult for users to know what they are looking at and trust it. The new rules aim to address that problem, while also raising questions about whether they will serve as a trust signal or add compliance burden, and whether they could lead to new forms of suspicion toward those who use AI for writing, from students to journalists and lawyers.