In August 2026, Anthropic began adding an invisible watermark to the text its newest models produce to meet transparency rules in the European Union's AI Act, and it applies to Claude users everywhere, not only in Europe. The watermark can't be seen, doesn't contain any special characters, and remains in the text when you copy and paste it. It won't affect the quality of the outputs.
If you've ever polished a resume with AI, you might have wondered if a recruiter could run the text through a detector and prove a machine helped write it. Though we won't know for sure until Anthropic's watermark checker tool is released, a resume is likely one of the hardest places for this watermark to be detected, and the reason has more to do with how a resume is written than with the watermark itself.
The signal hides in which word gets picked
When Claude writes text, it doesn't pick each word at random. It looks at several words that would all fit in the context, then uses a secret key and the words already on the page to choose one of them. Google DeepMind, whose method Anthropic uses, calls this "tournament sampling." Whenever a word is about to be produced, think of the model having a bucket of potential words to choose from (synonyms, for example), and a winner is selected based on a secret key.
All of those potential candidates are words that Claude might have used anyway, so Anthropic argues that the output reads the same as it would without the watermark. For example, the watermark won't push the output towards strange sounding words; otherwise, the watermark could be detected easily and removed. "Overcast might be selected in one sentence, grey in the next." Look at any single word and nothing about it seems unusual. The mark is the overall pattern across thousands of these small choices, and only someone with the key can see it.
Degrees of freedom set the strength
The watermark can only really be detected in cases where the model had a real choice to make, when several words would all work. But in cases where only one word fits, where a human would likely make the same exact word choice with near certainty, then there's nothing really to mark.
Consequently, length and topic have a large impact on the result. In a long stretch of regular writing, where a different choice could have been made for almost every word, then the mark is strong and easy to detect. In a short, factual passage, it is likely weak. Anthropic says that the mark is "sparser on factual passages where there are fewer choices," and it "doesn't work well on small samples." DeepMind gives its own examples, like naming a capital city or reciting a memorized poem, where the next word is almost fixed.
Placing a watermark in resumes is likely challenging
One caveat before going further. Anthropic has not released its detection tool yet, so this is reasoning based on the published method, not a tested result.
Based on the research, in order to detect the watermark, the text needed to have enough room to choose between different words, and be long enough to test several different words. A resume might be short on both.
Most of a resume is a list of facts. Names, employers, titles, dates, degrees, certifications, the tools someone has used. The model never had a choice between words when stating a proper noun, as "Disney" or "JavaScript" have no synonyms. The name of your team cannot be changed.
Other parts still follow strict conventions. A resume bullet has a rigid shape, a strong action verb, then a task, then a number. Led a team of eight, managed a regional budget, grew revenue 19%. The format narrows the word choice long before AI is involved.
Tailoring narrows it further. People or AI tools copy the exact wording of the job posting so the applicant tracking software counts them as a match, which forces the word choice toward a small fixed set of keywords.
A resume is short. Even where a real choice exists in a text bullet, there might be too few words for the signal to be detectable.
Light editing keeps the mark, a full rewrite removes it
The obvious next question is whether you can just edit the watermark out. According to DeepMind, the mark survives "cropping pieces of text, modifying a few words and mild paraphrasing." Fixing typos, swapping a few words for synonyms, tightening a sentence, none of it clears the signal. Only "thorough rewriting or translation" brings the confidence down to almost nothing, and by then you have basically rewritten the whole thing yourself.
Within hours of the announcement, Wired reported that coders had posted workarounds, mostly running Claude's text through a different AI model to reword it. It is unclear how effective these tools are until an official detection tool has been published.
Some models have no watermark
Watermarks are present in Claude and Gemini, which both use the same method, while they're not present in the many open models that anyone can download and run. Bots that submit hundreds of applications will likely avoid models with a watermark. So the applicant who used a mainstream tool like Claude to polish their resume might have traces of the watermark, while a bot flooding a job posting with near-identical applications might not have any watermark detected.
Given that the detection tool is not yet public, we've been unable to run tests on the watermark ourselves. Anthropic has shared that one is forthcoming, telling Wired, "We also plan to ship a text-detection API so users can do more of this themselves."
Works Cited
Anthropic. "How Claude's text watermarking works." Anthropic, 2026. https://www.anthropic.com/news/claude-text-watermark
Google DeepMind. "Watermarking AI-generated text and video with SynthID." Google DeepMind, 2024. https://deepmind.google/discover/blog/watermarking-ai-generated-text-and-video-with-synthid/
Dathathri, Sumanth, et al. "Scalable watermarking for identifying large language model outputs." Nature, 2024. https://www.nature.com/articles/s41586-024-08025-4
European Union. "Article 50, Transparency Obligations for Providers and Deployers of Certain AI Systems." EU AI Act, 2024. https://artificialintelligenceact.eu/article/50/
"Coders Say They Already Found Workarounds to Claude's Invisible Watermarks." Wired, 2026. https://www.wired.com/story/coders-say-they-already-found-workarounds-to-claudes-invisible-watermarks/