Latest online relationship
Anthropic, the artificial intelligence (“AI”) company which builds and distributes the popular Claude series of large language models, just set a potential trap which could affect millions of users, and most of them have no idea it’s there. The California-based company confirmed this month that Claude will weave a permanent, detectable watermark into its writing, a hidden fingerprint that survives copying, pasting, and even light editing. Nobody clicked “agree” to this. Nobody can turn it off.
Anthropic says the watermark is required under the new EU AI Act, a law designed to help citizens recognize when they are interacting with an AI system or exposed to AI-generated content. A separate set of voluntary guidelines that regulators treat as a safe harbor sets a rough exemption: AI-written passages shorter than about 150 words don’t need to be watermarked, because detection tools need enough words to distinguish an AI-written pattern from ordinary human writing that happens to overlap by chance. However, Anthropic chose to watermark everything Claude writes, even a single sentence. Anthropic also applied this approach worldwide, not just to users in Europe, likely sweeping in the everyday documents millions of people produce with AI help: academic papers, blog posts, school essays, cover letters, work emails, marketing copy, etc.
This overreach by Anthropic could cause real harm. Take a writer who brainstorms ideas with Claude, then spends extensive time rewriting key words and lines, adding original material, cutting what doesn’t work, and rearranging the structure. The finished post is essentially theirs. But because Anthropic watermarks text of any length instead of only longer passages per the EU guidelines, even the writer’s editing on their own screen may still leave enough leftover similarity for a watermark detection tool to flag the result as Claude-generated.
This could get worse for someone who edits directly on the Claude platform, specifies what changes to make, reviews the suggested edits, approves them, and then copies that edited text into their draft. Even though such changes reflect the writer’s own decisions and judgment, a checking tool would only see that final wording and might recognize all of it as coming from Claude.
Because of Anthropic’s overinclusiveness, those black marks from the checking tool could then be unfairly weaponized against the user. In the case of a political article or post, an opponent could disparage the writer as “using AI to write the piece.” Likewise, a hiring manager would not need to understand how the tool works to quietly pass on a candidate. A teacher would not need certainty to accuse a student of using AI to do their schoolwork.
Anthropic isn’t alone in having to comply with the EU law. However, OpenAI, owner and builder of ChatGPT, while bound by the same law, has not yet adopted a standard for its watermarking, citing exactly the false-positive risk Anthropic ignored. Other companies may end up following the EU’s voluntary 150-word threshold. That means users worried about unfair flags may soon have somewhere else to go, with Anthropic handing its competitors a marketing pitch for free.
Politically, Anthropic’s decision can also be seen as another example of a big tech company appeasing a regulatory body without fully weighing the impact on everyday people. It follows other criticisms of the company specifically: Anthropic paid $1.5 billion in the past year to settle a lawsuit from authors who accused it of training Claude on millions of pirated books without permission or payment.
Millions of people use AI to draft, brainstorm, and polish their own ideas every day. Anthropic’s answer to that reality is a permanent, detectable watermark that could get triggered too easily. That’s not transparency. It’s a potential liability AI users never agreed to carry.
Photo by Diari La Veu, used under Creative Commons license. https://tinyurl.com/4nu3hn58
