Anthropic adds invisible watermark to Claude Fable 5.1 and Mythos 5.1 outputs

Invisible watermark becomes default on new Claude text

1, and the launch statement said the company added a watermark — described as a numerical way of determining the likelihood that Claude was involved in writing a piece of text — to the outputs of the models. A help-centre article explained that Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch, and that new models will mark AI-generated content from day one. Independent coverage characterised the change as Anthropic embedding invisible watermarks in texts generated by its new Claude models so that AI authorship becomes machine-readable.

For practitioners, every paragraph produced by these models now carries a statistical signal that an automated checker can pick up, which changes how organisations document AI assistance in editorial workflows.

A privacy-first tool begins stripping AI watermarks

Within hours of the launch, an open-source project positioned itself as a privacy-first app that strips AI watermarks from content you own. The product surfaces a counter-trend: as default watermarking moves from research papers to production, tooling that detects the signal — and tooling that erases it — becomes a category in its own right. Editors and publishers who pass third-party AI-detection checks now have to reckon with the possibility that competitors, freelancers or bad actors may actively scrub the same signal. For a workflow that uses AI assistance but still wants the finished prose to read as purely human-authored, the artefact left in the text is no longer just stylistic.

A rewrite vendor markets a no-watermark output path

One vendor launched Phrasly Ultra, a rewrite product that the company says returns rewrites in moments and adds no AI watermark to its output by default. 8% Pangram human score across 72 evaluations, presenting the figure as evidence that the rewritten text reads as human on a widely used detector. Marketed as a way to make AI writing sound human, the product sits at the intersection of detection and evasion: it is offered as a drafting aid, while its defaults are explicitly chosen to avoid the very machine-readable signal that Claude has just made standard.

Buyers comparing tools now face a configuration choice as much as a quality choice — watermark-on or watermark-off, and whose detector defines "human".

What practitioners should check next

Three concrete checks follow from the September 29, 2026 changes. 1, because the watermark is new and detector calibration is untested at this writing. Second, audit vendor contracts: any rewriting tool that claims no AI watermark in its output is making a verifiable claim that procurement and legal teams can test against the same Pangram-style scoring.

Third, document the marking policy in editorial guidelines, especially for EU-published material, since the help-centre article ties day-one marking to models released on or after August 2, 2026. A standalone utility to estimate Pangram-style human scores on a draft, or to flag known watermark patterns in pasted text, would answer the questions editors are likely to ask first.