Here's everything you need to know about Claude's watermarking feature...
When it landed
Anthropic confirmed on August 11, 2026 that Claude models launched on or after August 2 embed a machine-readable watermark in everything they generate.
The rollout is global and applies at the model level, so it's there no matter where the text came from: the API, the Claude app, Claude Code, Cowork, Tag, and Claude accessed through AWS, Google Cloud, or Microsoft Foundry. There's no opt-out. Anthropic claims older models will get this feature over the coming months.
The reason isn't academic integrity. It's the EU AI Act. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside roughly 190 other signatories, which requires providers to mark AI-generated text in machine-readable form. Anthropic applied it worldwide because it can't currently scope the feature by region.
How it works
There are no hidden characters, and you can't strip it by pasting into a plain text editor. It has nothing to do with em dashes.
The watermark lives in word choice. When a model writes, it picks one token (or word) at a time, and often several candidates are equally good: "cold and overcast" versus "cold and grey." Normally a random number breaks the tie. With watermarking, the tie is broken using a secret key plus the preceding words. The choices still look random, but anyone holding the key can check whether the sequence matches what Claude would have produced with it, and assign a probability.
Some consequences worth knowing: code is barely watermarked, because exact output leaves no room for the nudge (comments can carry it, the code itself mostly can't). Factual writing carries less signal for the same reason. Light proofreading of your own draft may not register at all, since nearly all the words are still yours. Translation carries it fully, because Claude chooses every word. And the mark contains no identifying information about you, your org, or your chats.
How the public can check content
For text, right now: they can't.
Anthropic has said a detection API is coming and usable by third parties, but as of August 20, 2026 it hasn't shipped and there's no announced date. Anthropic holds the key, so no outside service can read the mark. Any tool claiming to detect Claude's watermark today is guessing from writing style, which is a different and far less reliable method.
How about files? Claude isn't an image generator in the Midjourney sense, so this covers a narrower set of output, but it can produce SVGs, charts and graphics rendered by code it runs, and images you upload that it converts or edits. To any of those it attaches a cryptographically signed content credential using C2PA, the open standard camera makers and photo editors already use. Any C2PA-aware tool can read it today. FYI - ChatGPT and Gemini do the same thing.
About removal
Within days of the announcement, developers published tools to strip the mark, one of which has passed 14,000 GitHub stars. Here's how those tools work: because the algorithm rides on word choice, anything that replaces enough of those words degrades it. Anthropic concedes as much, noting that light editing probably won't remove the watermark but a complete rewrite would. Running text through a second model has a similar effect, though that model's own provider may then be obliged to mark the new output. The legal position is unsettled. Researchers who've looked at it say the AI Act's marking duties fall on providers, not the public, and that neither the Act nor Anthropic's terms clearly bar making or sharing removal tools. The risk sits elsewhere: Anthropic's usage policy prohibits passing off model output as human-made, so stripping a mark in order to misrepresent AI work as your own is still a violation.
One caveat
Anthropic is unusually clear about the limits, and they deserve repeating, because schools, employers, and platforms will treat this as proof when it isn't.
A detected watermark means Claude was likely involved at some point. It cannot distinguish "Claude wrote this" from "Claude heavily edited this." And the absence of a watermark proves nothing: short passages, heavy edits, paraphrasing, or output from an older model can all come back clean.
The practical takeaway isn't to panic. It's to be upfront and disclosure if you used AI for your text and to keep your own records: drafts, timestamps, version history.
The watermark is a clue. Treat it like one.
Disclosure: AI was used to assist with editing and refining this article. The ideas and final content are my own.
Sources: Anthropic's help center article "How Claude marks AI-generated content" and its August 14, 2026 post "How Claude's text watermark works."