Anthropic To Watermark Content So People Can Tell It Was Processed By Claude

Anthropic has signed onto the EU AI Act's Code of Practice on Transparency of AI-Generated Content and will embed invisible watermarks in text and signed provenance metadata in files produced by its Claude models, with support for all Claude models launched in the EU on or after August 2, 2026. The marking will apply worldwide across all Claude surfaces, including the API, Claude.ai, Claude Code, Cowork, and Tag, and will also extend to models shipped before that cutoff, though no timeline is given. Anthropic says the marks are not a verdict on content origin, as they only indicate Claude may have processed the content at some point.

As AI-generated content proliferates across the internet — it surpassed human-written content recently — companies are taking steps so that people can tell it apart from human content. Anthropic has become the latest to formalise this, announcing that it has signed onto the EU AI Act’s Code of Practice on Transparency of AI-Generated Content, and detailing how Claude will start marking the text and files it produces. The company has published a breakdown of what this means in practice, and the short version is that Claude output is about to start carrying invisible fingerprints. Any Claude model launched in the EU on or after August 2, 2026 will support machine-readable marking right from launch, covering both generated text and generated files. Anthropic says it is also working backwards to extend marking to models that shipped before that cutoff, though there’s no firm timeline attached to that yet. Two layers of marking Anthropic is using two different techniques depending on what Claude produces. For text, it’s embedding a watermark directly into the output itself at the point of generation. This isn’t a visible tag or a disclaimer line at the top of a response — it’s woven into the text in a way that doesn’t change what the response says or how it reads. Because the mark lives inside the text itself, it should travel along when the content is copied and pasted somewhere else, and Anthropic says it may survive some amount of editing too. Since the watermarking happens at the model level, it applies no matter which Claude product generated the text, whether that’s the API, the consumer app, Claude Code, Cowork, or Tag. For files, Anthropic is turning to signed provenance metadata rather than an embedded watermark. When Claude generates a supported file type, such as an SVG, PNG, or JPG, it will attach a metadata label following the Coalition for Content Provenance and Authenticity’s C2PA standard, an approach that has become something of an industry default for tracking content origin. OpenAI took a similar route with DALL-E 3 back in 2024, and TikTok has leaned on C2PA credentials to label AI video at scale, tagging over a billion clips by its own count. The C2PA label on a Claude-generated file does two things: it signals the file passed through Claude, and it lets you check whether the file has been tampered with since. Where this shows up The coverage is meant to be broad rather than confined to the EU. Anthropic says marking will apply to output from supported models everywhere Claude is offered, worldwide, and across every surface — the API, Claude.ai, Claude Code, Cowork, and Tag. It extends to cloud access too: watermarks will show up when supported Claude models are run through AWS, Google Cloud, or Microsoft Foundry, though Anthropic notes that signed file metadata may not carry over to every one of those platforms depending on what features each supports. Anthropic is also promising to help people actually detect these marks, which is arguably the part that matters more than the marking itself. A watermark nobody can check for is not particularly useful. The company says it will support users and third parties in detecting Claude’s marks, as the EU Code requires, and that more detailed technical documentation on how detection will work is still to come. What the marks won’t tell you Anthropic has been fairly upfront that this system has real limits — the watermark can’t be treated as a verdict. A detected mark only tells you that Claude may have processed the content at some point — it doesn’t confirm the full history of that content. Someone might use Claude purely to proofread or translate a paragraph they wrote themselves, and the output would still carry a Claude mark despite the ideas being entirely human. Marked content can also be edited, excerpted, or folded into other material after Claude generated it, at which point the mark is describing an intermediate step rather than the finished product a reader sees. The reverse problem exists too. The absence of a mark doesn’t mean something wasn’t produced by AI. Anthropic lists several ways a mark can simply not be there: the content might have come from a model that predates marking support, the text might have been edited or translated heavily enough to scrub the watermark, the passage might be too short for a reliable signal, a file’s metadata might have been stripped by something as routine as a screenshot or a format conversion, or the feature might have been used on a platform that doesn’t support that particular type of marking yet. The bigger picture Anthropic’s timing tracks the regulatory calendar rather than any competitive pressure. The EU AI Act’s Article 50 2 obliges providers of generative AI systems to make AI output identifiable through technical means, and August 2026 is when compliance becomes non-optional for models launched from that date forward. Anthropic has been shipping models at a rapid clip this year — Opus 4.8 https://officechai.com/ai/claude-opus-4-8-benchmarks/ , Fable 5 https://officechai.com/ai/61673/ , and now Opus 5 have all launched within the past few months — so the marking requirement effectively applies to nearly its entire current lineup going forward. Google has a similar watermark https://officechai.com/ai/google-now-lets-users-check-if-an-image-is-ai-generated-through-the-gemini-app/ for images generated by its image generation tools, and SynthId also works for some other formats. For businesses building on top of Claude, Anthropic has flagged that the obligation doesn’t stop at the model layer. Companies deploying Claude inside their own products are told to independently work out what Article 50 requires of them specifically, with Anthropic promising more technical guidance on marking and detection as it becomes available. The message is essentially that Anthropic is handling the plumbing, but anyone building a product on top of Claude still needs to figure out their own transparency obligations to end users.