How Claude Marks AI-Generated Content
Anthropic will embed watermarks in the text and files generated by future models it launches in the EU, as part of its effort to comply with content and transparency rules in the bloc’s AI Act.
[…]
The move may further amplify the appeal of open weight models and alienate Claude customers, who don’t necessarily want consumers of their AI-generated content to know its provenance.
[…]
Anthropic itself is already hedging about the utility of its marking method, noting that detected marks are not conclusive evidence that Claude produced the content and that the absence of marks cannot guarantee that AI wasn’t involved in the creation of a particular piece of content.
Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported.
[…]
Marks will apply to output from supported Claude models across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered, worldwide. Some platforms or features may not support certain marking types.
[…]
We’ll support users and other third parties to detect Claude’s marks, as the Code requires, and we’ll share details in forthcoming documentation.
Anthropic is claiming this is for compliance with an EU law but that “Marking will apply to output from supported models wherever Claude is offered, worldwide.”
This is infuriatingly opaque. I won’t speak to the “signed provenance metadata” they say they’ll be embedding in generated files like PDFs, SVGs, PNGs, and JPEGs. That’s important but it’s complex. Plain text is simple. A string of text is just one character followed by another. I presume that Claude is going to start “embedding” invisible characters/byte sequences between visible characters? What they’re claiming to do here seems impossible, frankly, and anything they attempt to embed invisibly is going to cause immediate problems.
[…]
If the “watermark” is not comprised of invisible characters but rather visible ones, how in the world does this jibe with their claim that it’s “imperceptible”? (This Reddit thread claims that’s how it will work — “It’s a form of steganography where the model subtly biases its word choice to create a statistical pattern that can be detected later.” How in the world can that be squared with “it doesn’t change the meaning, quality, or readability”?) And any sort of semantic detection like this is going to cause false-positive problems.
Previously:
- Hiding Vulnerabilities in Source Code
- Hiding Data in Emoji
- A Vatican-Sized Flag Mystery
- Unicode and Copying and Pasting Code
- Hiding Easter Eggs in Maps
- Genius Accuses Google of Copying Its Lyrics Data
- Fingerprinting Swift Code Using Spacecrypt
- Xerox Scanners and Photocopiers Randomly Alter Numbers
Update (2026-08-17): Anthropic provides more details:
Watermarking won’t be specific to Claude. As of August 2, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.
[…]
In internal testing, we’ve seen no impact of watermarking on the content, level of creativity, or readability of Claude’s text. In the SynthID-Text paper, which introduced the technique we use, Google DeepMind tested this impact by serving a model that used watermarking to a portion of their Gemini traffic and comparing thumbs-up and thumbs-down ratings. They found no statistically significant differences from the unwatermarked model. And in a controlled study, human raters comparing watermarked and unwatermarked answers side-by-side saw no difference in quality.
[…]
There are limitations to the effectiveness of watermarking. Using our key, one can only answer the question “What is the likelihood this was partly written by Claude?” It doesn’t confirm whether the text was human-written, and it can’t tell whether the text was written by a different AI (even if that other AI uses watermarking, it would have a different key; it might also use a different watermarking method altogether). Detecting a watermark also doesn’t work well on small samples, where there are fewer word choices and thus less information to go on. As a passage increases in length, confidence about Claude’s involvement increases too.
Watermarking is sparser on factual passages where there are fewer choices that can be made without decreasing the accuracy of the text. […] For the same reason, code—which in very many cases has to be exact—has generally less watermarking than some other forms of text.
John Gruber (Mastodon, Bluesky, Hacker News):
I initially guessed “invisible characters” not because I didn’t think of the semantic word-choice technique, but because I was a fool who took Anthropic at its word in their description of what they would do.
[…]
They say “imperceptible” and “doesn’t change the meaning, quality, or readability”. Their words. Not almost imperceptible. Not slightly changes the meaning, quality, or readability.
There seems to be a slight of hand here, where Google studied whether humans saw a difference in quality but then Anthropic claimed that there was no difference in meaning. Changing words will almost always change the meaning. Perhaps not in an important way, but a change is a change. Pretending otherwise is offensive to a writer. The response to this line of argument seems to be that LLMs already have randomness and that the marking changes will be within the level noise that’s already there. To me, this presents a pessimistic view of the capabilities of LLMs, contrary to Anthropic’s normal claims. The Google study was published in 2024 and didn’t test Claude. The models have changed a lot since then, but they are assuming that the precision hasn’t advanced enough to be affected, nor will it in the future. Or that when there is a tension between watermarking and quality they will pick the former. And perhaps you won’t be able to tell because everyone else will be watermarking, too.
But the very best description of the general idea behind the technique is an interactive essay by James Padolsey, “How AI Text Watermarking Works”.
[…]
This isn’t just about text one might generate with the intention of passing it off as their own natural work. This isn’t even about LLM proofreading of work written by hand. Anthropic is saying that all new Claude models are going to adulterate every single bit of text longer than 200 tokens (~150 words) they generate, including everything it presents to its users to read. So even in a private conversation between a user and Claude, which will never be read by anyone other than the user, Claude will begin making word choices in the name of marking its output in statistically predictable ways rather than maximizing clarity and precision.
[…]
Taken literally, compliant LLM terms of service must forbid users from rephrasing the output from models that comply with this regulation, because the word choices are the marks. But it’s not the European Union that is trying to impose their absurd, impractical, witch-hunt-fueling regulation on the entire world. That falls on Anthropic.
Complying with this, particularly with regard to text, is only going to create problems for honest users. Dishonest users attempting to pass off AI-generated text as their own writing (students, employees, whoever) will simply circumvent detection through non-compliant AI paraphrasing tools.
I don’t love that checking for a watermark requires submitting the text to the AI provider, and all you get back is an opaque result with no way to verify it.
James Padolsey (via John Gruber):
The move is Anthropic’s response to Article 50(2) of the EU AI Act, which requires providers to ensure that outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”.
[…]
The Act exempts standard editing, but many legitimate assistive uses require more substantial rewriting while leaving the ideas, judgment and responsibility with the human. The same thought that led to this law could have applied to calculators at the time of their inception, had their outputs revealed themselves through artefacts. Thankfully, a sum borne of the brain is treated no differently from one produced by a calculator. Likewise with spellcheckers. To make assistance suspect only once the tool becomes capable enough to compose a whole sentence is not a principled boundary. It is a moral premium placed on difficulty itself.
Anthropic has nevertheless chosen a blanket, model-level implementation that appears broader than the law’s minimum requirement. That may be convenient compliance engineering, but it discards distinctions the law expressly attempted to preserve. The result is a signal broad enough to implicate harmless and assistive use, yet fragile enough to be removed by a motivated person through substantial recomposition. It risks concentrating suspicion on ordinary and assistive users while remaining weakest against deliberate deception.
Update (2026-08-20): See also:
- Supplementary Information for the SynthID-Text paper
- Jason Anthony Guy
- Daniel Jalkut (Mastodon)
- John Gruber (Mastodon)
- Greg Titus
- Allie
Without getting into the weeds, I still think Anthropic’s claim of their being no change in the meaning, quality, or readability is not accurate, using the plain meanings of those terms. If you read the papers, there are ways of ensuring that the frequencies of individual tokens or sequences of them are not affected in aggregate. But this is a different claim, and Anthropic does not specify which tradeoffs they’ve chosen there. I’m willing to grant that they can make the effect very small.
At this point, I think the most interesting question is, what will be the effects of these watermarks?
AI watermarking is security theater. There is absolutely nothing to prevent people from removing any possible watermark in text, images, video or other data. There is no way to prevent false positives, or false negatives, in detectors, either. It’s all superficial nonsense.
I find it interesting to be able to test a particular passage of text, but what am I supposed to do with that information? Even without watermark removal, a negative for a given model doesn’t tell you whether another model was used. And I’m uncomfortable with convicting someone based on a watermark given that false positives are possible and independent verification is not.
I asked Gemini about the benefits, and it said:
Regulatory Compliance and Policy
- Legal adherence: Helps major tech platforms satisfy emerging mandates requiring automated systems to label machine-generated media.
- Ecosystem tracking: Gives platforms and enterprises a high-level telemetry signal to monitor the spread of synthetic content across public channels.
Triage and Risk Mitigation
- Mass moderation: Assists automated filters in flagging bulk, unvetted bot content or spam networks where statistical trends matter more than single-file perfection.
- Provenance auditing: Provides an extra layer of metadata for digital media archives, helping investigators spot obvious clusters of AI-generated assets.
In other words, it’s not so much about individual pieces of content.
Update (2026-08-24): John Gruber:
Also, more and more, I do think this has little to do with the EU regulation and everything to do with Anthropic and Google wanting to watermark the text that Claude and Gemini generate for their own purposes, to identify text to avoid training future models on to avoid model collapse/AI inbreeding.
See also: Accidental Tech Podcast.
