Anthropic just answered the questions everyone’s been asking about Claude’s new text watermarks. In a blog post published Friday, the company laid out how the watermarking works, whether editing can strip it, and what it means for code, according to TechCrunch AI. This follows an announcement earlier in the week that caught a lot of Claude users off guard, and TechCrunch AI reports the reaction has been loud.
What stands out here is that Anthropic isn’t doing this by choice. It’s complying with the EU AI Act’s Transparency Code, which requires AI companies to make AI-generated content identifiable. And Claude won’t be alone. Anthropic says other major model developers have signed the same Code of Practice and will roll out their own watermarks too.
What a watermark actually is
Here’s the mechanic. When Claude makes what Anthropic calls “low-stakes choices,” like picking between “overcast” and “grey” to describe the weather, it can bake a pattern into those choices. You can’t see it as a reader. But anyone holding the right key can detect it.
“Watermarking does not impact the quality of Claude’s output,” the company said, per TechCrunch AI. “To a reader, a watermarked response is indistinguishable from an unwatermarked one.”
Anthropic is using the SynthID-Text approach that Google DeepMind published in 2024, and it plans to release a detection API so the watermark can be checked.
One important distinction: this is not the same as AI-detection tools from companies like Pangram, which hunt for stylistic “tells” in writing. Anthropic says spotting those patterns is “fundamentally different from checking for a watermark.” A watermark is a deliberate signal. A tell is a guess.
Can you just edit it out?
This is the question users care about most, and Anthropic’s answer is honest. Yes, it’s possible, but not easy.
- Light editing probably won’t remove the watermark completely.
- A complete rewrite where every word is replaced will strip it.
And Anthropic makes a fair point about that second case: “it’s arguable whether the text can any longer be described as AI-generated.” If you’ve swapped every word yourself, how much of it is still Claude’s?
For text that Claude only proofreads or edits, detectability depends on length and how heavily Claude touched it. Light edits mean “nearly all the words” came from the human author, so there’s “very little (if anything) for the watermark to attach to.”
What about code?
Good news for developers. Code should carry far less of a watermark than regular text. The reason is simple: working code doesn’t leave room for arbitrary word swaps. The model has to produce something that runs, so it can’t freely pick between equally valid options.
Where there is wiggle room, like comments inside the code, the watermark can show up. But Anthropic says it’ll have “a negligible effect on the actual code produced.”
Why this matters
The backlash tells you something. On Reddit, one poster called the move a conspiracy against innocent Claude users, while another argued, “The only reason you wouldn’t want this is to lie to people.” Business Insider reports that dozens of X users claimed to cancel their subscriptions over it.
That gap in reactions is the real story. Watermarking is arriving because European regulation demands it, not because AI companies volunteered. So the trade-off is now out in the open: provenance and accountability on one side, and users who feel surveilled on the other.
For practitioners, a few things are worth watching:
- Detection is coming with the watermark. Anthropic’s planned API means anyone with access can verify Claude-generated text, not just Anthropic.
- This is an industry standard, not a Claude quirk. Other signatories to the Code of Practice will ship their own versions, so expect watermarking to become the default across major models.
- Your code is mostly safe. If you’re using Claude for engineering work, the watermark barely registers.
Watermarking is now part of how frontier AI ships in a regulated world. Anthropic is just first to explain the details in plain language, and the rest of the field is right behind it. You can read the full breakdown at the original source.