Anthropic publishes a watermarking FAQ four days into the user backlash

Anthropic answered the watermark panic, and the answer amounts to: the internet spent the week arguing about the wrong watermark.

What the FAQ says

Four days into the drama, Anthropic published an official FAQ on August 14, 2026. The company’s key claims: watermarking exists to comply with the EU AI Act, and other major model developers signed the same Code of Practice, so they will be implementing watermarking too. Nothing is added to the text and there are no hidden characters. The difference between watermarked and un-watermarked text is not distinguishable to readers. Watermarking consumes no extra tokens and costs nothing. And watermarks cannot be traced to a specific person, organization or chat.

The collision with the eraser

The timing gives the FAQ an unusual edge. Earlier the same day, an open-source “watermark eraser” that strips invisible Unicode characters from AI text passed 6,000 GitHub stars. Anthropic’s FAQ says its watermark contains no hidden characters at all.

If both statements are accurate, the week’s most viral removal tool is cleaning marks this watermark never carried, while the statistical mark, the one embedded in word choice itself, stays exactly where it was.

Where the saga stands

The sequence has run at internet speed: rollout became public on August 11, the user revolt peaked on August 12, the eraser went viral on the morning of August 14, and the FAQ landed that evening. The remaining tension is not technical but social: “cannot be traced to a specific person” addresses the surveillance fear, but the workplace fear, an employer running a detector over an employee’s writing, is the scenario users were angriest about, and detection is exactly what a watermark is for.