What changed
Anthropic named the method behind Claude’s text watermark in a post dated 14 August 2026 1. It is a version of the SynthID-Text approach Google DeepMind published in a Nature paper in 2024, in a family going back to a proposal by Scott Aaronson in 2022, whose stated design principle is that the watermark only changes the source of the randomness used to pick among words 1. Search Engine Journal, reading the obligations attached to Claude’s text watermarking on 11 August 2026, notes that published text under qualifying human review or editorial control can be exempt from the disclosure requirement 2.
The mark and the duty are two different objects
Search Engine Journal states it outright: a Claude mark can appear on copy that its publisher has no obligation to label 2. I think anyone running a content pipeline should sit with that sentence. The exemption turns on who holds editorial responsibility, a fact about people and process 2. The mark attaches to token choices, a fact about generation 1. Neither can answer for the other.
One consequence I would add: with no randomness there is no randomness source to change. Output your stack produces with the sampler pinned leaves nothing to substitute, so marking looks different for template-shaped copy than for free-form prose. That is my reading, and it is not a measurement.
Impact on your team
The decision sits at generation time: what your pipeline records when it writes the text. If you push Claude output into a CMS or a customer-facing template, store the editorial fact as you write: who read the copy, under what process, and whether that reading is the qualifying kind Search Engine Journal describes 2. It does not come back from the bytes later. Two moves I would refuse: treating a mark as proof its publisher owed a label, when the exemptions run the other way 2; and treating the sampler as an implementation detail, when it has just joined your provenance surface.