You paste text generated by Claude into a professional document, and a question crosses your mind: can anyone prove you didn't write it?
Since Anthropic announced a watermarking scheme for its models, the answer is yes, to some extent. The Claude watermark is an invisible signal embedded in generated text.
It makes it possible to calculate the probability that Claude produced it, without adding a single hidden character or changing the meaning of a single sentence.
This scheme applies to Claude models launched after August 2, 2026, as part of the European Code of Practice on transparency for AI-generated content.
Here's how it actually works, what does and doesn't get marked, and what it means if you're producing content solo, without a legal team on hand.
How the watermark slips in without adding anything to the text
A model like Claude doesn't generate a whole sentence at once. It picks one word after another, choosing at each step among a list of plausible candidates.
After "The weather today was cold and…", the words overcast and grey are just as valid as each other. The final choice doesn't change the meaning of the sentence at all.
In cases like this, the choice is settled by a random draw, a bit like an invisible dice roll made for every word in the text.
The watermark doesn't touch that principle: it never forces Claude to pick a word it wouldn't have chosen naturally. It simply swaps out the source of the randomness.
Instead of a purely arbitrary draw, Anthropic uses a secret key combined with the preceding words to determine the choice.
The result is still a word Claude could have picked anyway. But the whole sequence becomes verifiable to anyone who holds the key.
That's what makes it possible, after the fact, to calculate a probability that a given text came from Claude, without anything having been added or hidden inside it.
Anthropic is reusing a method published by Google DeepMind in a 2024 scientific paper, called SynthID-Text. It builds on a proposal by researcher Scott Aaronson from 2022.
So this isn't an isolated Anthropic invention, but an approach already tested at scale on other models, including Gemini.
No invisible character is inserted, contrary to what some tools suggest when they promise to "clean" Claude's text.
These tools generally target zero-width spaces: invisible unicode characters sometimes used by other detection systems to mark a text.
Claude's watermark doesn't rely on that mechanism. It lives in the word choices themselves, not in symbols added afterward.
What this actually changes for text quality
Anthropic states it observed no effect on the creativity or readability of Claude's responses in its internal testing.
Google DeepMind had reached the same conclusion earlier, by serving a watermarked model to a portion of Gemini's traffic.
Comparing thumbs-up and thumbs-down ratings on watermarked versus unwatermarked responses, the team found no statistically significant difference in satisfaction.
Watermarking also doesn't produce any extra tokens. A token is the unit of text, often a word or part of a word, that a language model bills for.
So there's no added cost or slowdown for you, whether you use Claude through the web interface or through the API.
Why code and factual text barely get watermarked
The watermark needs arbitrary choices between several equivalent phrasings to take hold.
Code offers almost none: a line that computes a sum has only one correct output. A mis-named variable breaks the program.
The same logic applies to purely factual passages. If Claude writes "Isaac Newton's most famous work was called Principia…", there's only one correct word to finish the sentence.
No room for a random draw, so no room for the watermark.
In code, the signal can still show up in comments, where several phrasings remain equivalent. But its effect on the code itself stays negligible.
Proofreading vs. translation: two radically different cases
If you ask Claude to fix the spelling and punctuation of a text you wrote yourself, almost all the words stay yours.
The watermark then has almost nothing to attach to. Depending on how extensive the corrections are, Claude's involvement can remain undetectable.
Translation works the opposite way. Claude chooses every single word of the translated text, since it has to rebuild the whole sentence in another language.
The result is therefore fully watermarked, and easily identifiable as generated content, even if the source text came entirely from you.
Images and files: a signature, not a watermark
Files produced by Claude, such as a .png, a .jpg, or an .svg, don't get the same treatment as text.
They carry a cryptographically signed note in their metadata, using the open C2PA standard (Coalition for Content Provenance and Authenticity).
It's a traceability standard already used by camera manufacturers and photo-editing software, to indicate where an image comes from.
This isn't a watermark in the strict sense: nothing is modified or hidden in the image's pixels.
It's a signature that any tool compatible with the standard can verify, including the one Anthropic plans to provide.
Like any metadata, it doesn't necessarily survive a format conversion or a simple screenshot, even though the visual content itself stays unchanged.
The watermark never traces back to you
Neither the watermark nor the key used to read it contains any information about the user's identity, their organization, or their conversation.
The marking applies only to the words chosen by the model. The key answers a single question: how likely is it that Claude was involved in this text?
It never says whether Claude wrote the text from scratch or simply touched it up.
That's an important limitation if you were counting on this mechanism to prove the exact origin of content in a dispute.
Can the watermark be removed?
Partially. A light edit, two synonyms swapped or three commas moved, isn't enough to make the statistical signal disappear.
It takes a deep rewrite, where nearly every word changes, to seriously weaken it.
Anthropic itself notes that a text where every word has been replaced raises a deeper question: can it still be called AI-generated content at that point?
Detection also gets less reliable on short text, since there aren't enough word choices to establish a solid probability.
Some online tools market themselves as "watermark removers" and promise to clean invisible characters out of text.
They're solving a different problem: the zero-width spaces or em dashes that other detection methods sometimes rely on, not the statistical mechanism Claude uses.
Getting around it requires a full manual rewrite, not a button click.
How to check if a text was generated by Claude
Anthropic plans to offer a detection API, with the exact details still to be worked out.
It will function differently from the AI text detectors available on the market today, like Pangram, Originality.ai, or Winston AI, which don't have access to Anthropic's key.
These independent detectors rely on writing tics typical of language models.
For example the "it's not X, it's Y" construction, or the frequent use of certain adverbs (like "quietly").
The result is always a probability, never a certainty. None of these tools can tell text Claude wrote from text Claude merely reworked.
What this means for you in practice, even solo
You can no longer rely on an older model to durably dodge the watermark. Claude versions from before August 2, 2026 still escape it.
But Anthropic has a transition period and is working to cover them in the coming months. Switching to an older model only buys you time, not a solution.
You also can't get around it with a VPN. The marking operates at the model level itself.
It applies globally at launch, with no geographic distinction. Connecting from outside the European Union changes nothing about the text Claude produces.
You should treat translation and pure generation as higher detection risk, unlike light proofreading.
If you publish a text Claude translated in full, assume it's identifiable as AI content. If you only ask for a grammar fix on your own writing, the risk stays much lower.
FAQ
What is the Claude watermark?
The Claude watermark is an invisible statistical signal, built into the word choices made during text generation.
It doesn't change the meaning of the content and adds no hidden character. It lets anyone holding Anthropic's key calculate a probability that Claude took part in writing a given text.
Does the Claude watermark slow down or add cost to using the model?
No. Watermarking produces no extra tokens and showed no effect on generation speed or cost in Anthropic's internal testing.
Claude's pricing and latency stay identical whether the watermark is active or not.
Can the watermark be removed from Claude-generated text?
A light edit isn't enough to erase the signal. It takes a deep rewrite, replacing nearly every word, to seriously weaken it.
Tools that promise to strip invisible characters are solving a different problem and have no effect on this statistical mechanism.
Can the Claude watermark identify who used the AI?
No. Neither the watermark nor its reading key contains information that could identify a specific user, organization, or conversation.
The mechanism only answers the question of the probability that Claude was involved in the text, not the question of its human author.
Is code generated by Claude watermarked like text is?
Very little. The watermark relies on equivalent word choices, which barely exist in code where an exact output is required for the program to work.
The signal can appear in comments, but its effect on the code itself stays negligible.
