Anthropic has begun weaving invisible watermarks into the words its AI produces, a worldwide change that took effect on August 2 and arrives with no option to switch it off, a notable step for a technology whose output has been almost impossible to distinguish from human writing.
The watermark works by nudging Claude’s word choices in a subtle, patterned way, small enough that a reader would never notice it in a single paragraph but consistent enough to become statistically detectable across a longer stretch of text. Because the pattern lives in the choice of words rather than in any hidden character or file tag, it travels with the writing when someone copies and pastes it, which is precisely what a mark meant to follow content out into the world needs to do.
What the mark can honestly tell you is narrower than it first sounds, since Anthropic is clear that a detected watermark means Claude may have processed the text, not that Claude wrote it. A person who runs their own draft through the model to proofread it, translate it, or shorten it can end up with a watermarked result even though every idea and most of the words are theirs, so the signal points to involvement rather than authorship, an important distinction for anyone tempted to treat detection as proof of cheating.
For files rather than plain text, Anthropic is leaning on a different tool, attaching provenance metadata in the widely used C2PA standard that records the model’s involvement and helps reveal later tampering, though the company acknowledges this marking is more fragile, capable of being stripped away by converting a file to another format, re saving it, or simply taking a screenshot. The text watermark is sturdier but not permanent either, since heavy rewriting or paraphrasing will eventually scrub it out.
The reason for all of this sits in Brussels rather than in any sudden change of heart, because Article 50 of the European Union’s AI Act now requires that machine generated content be marked in a machine readable way, with penalties that can reach 15 million euros or 3 percent of a company’s global annual turnover. Rather than build one system for Europe and another for everyone else, Anthropic chose to apply the watermark to Claude’s output worldwide, a decision that has drawn its share of pushback online from users who dislike being marked by default. In the interest of disclosure, Entrelligence uses Anthropic’s Claude among the tools in its newsroom, and this policy applies to those tools, so we held this article to our usual sourcing and review. The move is a real step toward transparency in a field that badly needs it, and at the same time a reminder that watermarking is a soft signal rather than a hard proof, useful for tracing where text came from but easily defeated by anyone determined to hide it.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →