{"slug": "the-end-of-undetectable-ai-text-claudes-new-watermark-explained", "title": "The End of Undetectable AI Text? Claude’s New Watermark Explained", "summary": "Anthropic has signed the EU AI Act's Code of Practice on Transparency of AI-Generated Content and announced an 'imperceptible watermark' for Claude text outputs, designed to survive copy-paste and some editing. The company has not yet disclosed the technical details, but experts speculate it may use statistical token watermarking similar to Google DeepMind's SynthID Text. The move aims to make AI-generated content machine-detectable, with exceptions for human-reviewed content.", "body_md": "For the past few hours, the whole world, or at least my LinkedIn feed, has been talking about one thing: Anthropic signed the **EU AI Act’s Code of Practice on Transparency of AI-Generated Content**.\n\nIs this the end of generating funny cat stories, LinkedIn posts, and boring docs? Are we doomed? Will we have to write all those boring things manually, like animals, because otherwise some mysterious detector will expose us? 😅\n\nI just couldn't resist writing about this! The internet has already managed to produce a surprising number of myths around the topic, so let's quickly go through what we actually know for sure, which, by the way, is pretty easy to find on Claude's official pages, where the documentation is more pleasant to read than many industry articles.\n\nAnd as you'll see, there's really not that much to be afraid of. Also... Anthropic isn't even the first company to do this.\n\nOf course not!\n\nIf you're an EU citizen like me, you can still ask an AI to write a reply to a boring email or generate a ridiculous poem, and you won't go to jail for it. 😉\n\nThe goal isn't to ban AI. It's about making AI-generated and manipulated content machine-detectable and, in specific contexts, transparent to users — especially when we're talking about things like deepfakes or content concerning matters of public interest.\n\nAnd there is an important nuance here: the AI Act also provides an exception for certain AI-generated or manipulated text where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication.\n\nSo no, the EU has not declared war on your ChatGPT-generated emails.\n\nHere's what we know for sure today:\n\nAnthropic calls it an **\"imperceptible watermark\" embedded directly into the text**.\n\nAccording to Anthropic, it survives copy-paste and can even survive some amount of editing. It's implemented at the model level rather than being something the Claude frontend simply adds afterward.\n\nBut here's the important part: **Anthropic still hasn't told us exactly how it works!**\n\nThe company says more technical documentation is coming, so... expect a follow-up from me when it arrives. 😀\n\nWe can make an educated guess that it might be some form of **statistical token watermarking**. In such systems, we're not talking about weird spaces or invisible Unicode characters that could be removed in a second. Instead, the model can slightly and statistically prefer certain tokens during generation.\n\nIn a single sentence, that kind of signal might be impossible or unreliable to detect. Across a longer piece of text, however, statistics can start revealing the pattern.\n\nSo, very roughly, you could imagine something like this (**if Anthropic has implemented its watermark this way)**:\n\n**Normal generation**\n\n```\nP(next_token | context)\n```\n\nvs.\n\n**Watermarked generation**\n\n```\nP(next_token | context) + secret statistical bias\n```\n\nBut again: **Anthropic has NOT confirmed yet that this is how Claude's watermark works.**\n\nThis is simply an educated guess based on existing text-watermarking techniques. 🙂\n\nNope!\n\nGoogle has actually been doing this for years.\n\nGoogle DeepMind introduced **SynthID Text** back in 2024, and it works along the general lines I described above: the watermark is introduced during token generation, creating a statistical signal that can later be detected.\n\nOpenAI, meanwhile, has said that it has developed text-watermarking technology, although while it already uses provenance or watermarking technologies for things like images, video, and audio, it has not publicly rolled out comparable watermarking for ordinary GPT text responses.\n\nSo Anthropic definitely didn't invent the idea of watermarking LLM output.\n\nWe don't really know yet.\n\nFor supported files, the situation is somewhat simpler because Anthropic uses provenance metadata, such as C2PA credentials, that can be inspected with appropriate verification tools.\n\nText is more interesting.\n\nAnthropic says it plans to make watermark detection available to **third parties and researchers**, but we don't yet know exactly who those third parties will be, what access will look like, or whether regular users will eventually get a simple public detector.\n\nSo GPTZero and similar garbage may still have a job for a while.\n\nWill everything finally become crystal clear? Will every LLM-generated article suddenly get a giant red **CHEATER!** warning next to it? xD Of course not!\n\nAnthropic itself openly talks about the limitations of this technology. And we've discussed variations of this problem thousands of times here on DEV already.\n\nFirst of all, the presence of a watermark does **not** necessarily mean that the author didn't write the text themselves.\n\nMaybe they wrote the entire thing and then asked Claude to edit it, fix grammar mistakes, polish the wording, or translate it into another language. The resulting output could contain a watermark even though the actual ideas, and potentially most of the original text, came from the human author.\n\nSo yes, the final article could still be chaotic, unnecessarily long, and full of a million digressions, exactly like my posts here.\n\nAnthropic explicitly points out that the opposite can happen too.\n\nSomeone can take fully AI-generated slop and heavily edit it, paraphrase it, translate it, mix it with other content, or otherwise transform it enough that the watermark may no longer be detectable.\n\nSo:\n\n**watermark detected ≠ AI wrote everything**\n\nand\n\n**no watermark detected ≠ human wrote everything**\n\nThat's a pretty important distinction.\n\nWith supported files, provenance can be easier to verify because we can use embedded metadata and content credentials. But even that isn't bulletproof.\n\nWhat happens if someone takes a screenshot of an AI-generated image, for example Metadata doesn't magically survive every possible transformation.\n\nAnd that's exactly why provenance is useful evidence, but not some perfect universal AI detector.\n\nThis is probably the part I'm most curious about. How exactly are they planning to handle watermarking in generated code, where the model has significantly less freedom in choosing equivalent tokens?\n\nIn natural language, there can be dozens of perfectly valid ways to express essentially the same thought. But if we have something like:\n\n``` js\nconst result = await fetchData();\n```\n\nthen how the hell do we add a statistical watermark to that?\n\nSure, there are still choices: variable names, formatting, and so on. But the space of valid substitutions is obviously much more constrained than it is in natural language, and formatters, refactoring, minification, or even small manual edits could make the problem even more interesting.\n\nClaude Code is included in Anthropic's marking rollout, but Anthropic hasn't yet published enough technical detail for us to know exactly how the text watermark behaves with generated source code.\n\nAnd I really want to know how they solved this one.\n\nAs you can see, Claude watermarking generated text is actually a logical and important, although obviously still imperfec, move.\n\nIt probably won't change our everyday lives very much.\n\nSure, the flood of AI slop across social media is annoying. Unfortunately, I have a feeling this watermark isn't going to magically solve that problem either. 😐\n\nWhat interests me much more is **how this watermark is actually implemented**.\n\nAnthropic, I'm waiting for the docs!", "url": "https://wpnews.pro/news/the-end-of-undetectable-ai-text-claudes-new-watermark-explained", "canonical_source": "https://dev.to/sylwia-lask/the-end-of-undetectable-ai-text-claudes-new-watermark-explained-45g2", "published_at": "2026-08-11 22:33:57+00:00", "updated_at": "2026-08-11 22:45:47.318968+00:00", "lang": "en", "topics": ["ai-policy", "ai-safety", "large-language-models", "ai-products"], "entities": ["Anthropic", "Claude", "EU AI Act", "Google DeepMind", "SynthID Text", "OpenAI"], "alternates": {"html": "https://wpnews.pro/news/the-end-of-undetectable-ai-text-claudes-new-watermark-explained", "markdown": "https://wpnews.pro/news/the-end-of-undetectable-ai-text-claudes-new-watermark-explained.md", "text": "https://wpnews.pro/news/the-end-of-undetectable-ai-text-claudes-new-watermark-explained.txt", "jsonld": "https://wpnews.pro/news/the-end-of-undetectable-ai-text-claudes-new-watermark-explained.jsonld"}}