When an O'Reilly instructor faced a live Q&A about Claude Agents, the audience had one overwhelming question: how AI watermarking works and what it means for them. The interest surfaced shortly after Anthropic revealed plans to embed invisible markers in all text outputs from its newest models.
Anthropic’s announcement and EU compliance
In August 2026, Anthropic confirmed that future Claude models will generate text containing a watermark. The purpose is to determine the likelihood that Claude contributed to writing the text. Anthropic along with several other major AI providers are making this change to comply with the EU AI Act. A second quote from the same announcement makes clear that watermarking is not unique to Claude: as of August 2, 2026, the EU requires any AI provider serving its market to mark AI-generated content. Other major model developers have signed the same Code of Practice and will implement their own watermarks.
The specific Claude models that include AI text watermarks in all generated outputs are:
- Claude Mythos 5.1
- Claude Fable 5.1
- Claude Fable 5
- Claude Opus 5.5
- Claude Opus 5
- Claude Opus 4.8
- Claude Sonnet 5.5
- Claude Sonnet 5
- Claude Haiku 5.5
It does not matter whether a user accesses Claude through Desktop, Code, or the API — every output carries the watermark. Consequently, any code, email, blog post, article, newsletter, or white paper written with Claude will bear the marker, and any AI watermark detector can flag the content as AI-generated.
How the watermark actually works
Anthropic’s watermarking technology is SynthID-Text, an open source project from Google DeepMind. SynthID-Text is a form of steganography, not cryptography. Cryptography, familiar from HTTPS websites, encrypts content so it becomes unintelligible without a decryption key. Steganography hides information in plain sight; to an observer the text looks ordinary, but someone who knows where to look can detect the hidden signal.
In practice, the watermark embeds itself through word choice. The model selects among near-equivalent words — for instance, choosing between "big" and "large" — without adding hidden characters, without inserting extra tokens, and without embedding any user-specific data. An AI watermark detector can determine whether text is AI-generated, but it cannot identify which Claude account produced it.
The technical method is tournament sampling. At every position, the model draws several candidate tokens and runs them through a knockout-style bracket. A pseudorandom score, called a g-value, decides the winner. That g-value comes from hashing the previous few tokens together with a secret key. Across a text of hundreds of words, the output subtly favors tokens with high g-values in a pattern only the key holder can perceive. A reader glancing at the text would never notice the difference. Google tested this approach on roughly 20 million Gemini responses and found the thumbs-up rate differed by 0.01%, with the watermarked model holding a slight edge.
Industry-wide adoption and regulations
Anthropic is not alone. OpenAI has announced it will add invisible watermarks to eligible ChatGPT and Codex text output in the European Union. OpenAI calls its technology textGrain, drawing an analogy to writing on a grain of rice. According to OpenAI’s own tests, their watermark detector achieves 95% accuracy on passages containing 400 tokens — roughly a few paragraphs of text.
Because the EU rule covers more than text — it extends to images, audio, and video — OpenAI has also launched a public website where anyone can check whether non-text media carries a watermark. The same EU regulation is mirrored in California through SB-942, the AI Transparency Act, which also mandates watermarks on AI-generated content.
Google had already complied with the EU AI Act, equipping its models with the SynthID Text Watermark as early as 2024. Google later released SynthID.com, a public portal where users can check non-text media for watermarks from multiple providers including OpenAI, Nvidia, and Kakao, in addition to Google’s own outputs.
Why watermarking matters
A consultant once witnessed a real-world case where scammers, having harvested employee names from LinkedIn, impersonated a CEO via email and demanded bank transfers from personal accounts. The fraud failed only because the attackers used external email domains. But voice cloning technology already requires only a few voice samples, and a more convincing spear-phishing attack could use AI to clone a CEO’s voice over the phone. In that scenario employees might obey instructions they believe come from leadership, with potentially severe financial consequences.
Embedding watermarks in AI-generated content — whether text, images, audio, or video — gives society a tool to assess authenticity. Bad actors can exploit AI for sophisticated attacks, but watermarks help identify synthetic material.
All major large language models now include watermarking, and regulations ensure it is not going away. Developers and teams building with these models should expect every output to carry a marker, and should factor detection capabilities into workflows that publish or distribute AI-generated material.