Anthropic explains Claude's invisible text watermark as Gemini's visible image marks go optional

What Claude's invisible watermark actually is

The most concrete change is the explanation of how Claude marks its text. Coverage describes the technique as an invisible statistical signal built into the model's sampling step rather than a visible stamp appended after the fact, with the underlying research identified as Google DeepMind's SynthID-Text published in Nature. The mechanism, called Tournament Sampling, is said to construct a balanced binary tree that lets the system preserve text quality while still producing a detectable signal, and Anthropic's write-up reportedly addresses how that statistical signature can be erased.

For practitioners, the workflow change is simple: any text now coming out of Claude carries a forensic mark by default, which means downstream pipelines that re-tokenise, paraphrase, or translate the output can attenuate or remove it before it reaches an end user.

Independent testing finds no watermark in newer Claude and GPT models

An independent plain-language test reported on 2026-08-24 looked for the SynthID signal in current frontier outputs and concluded that no watermark was present in Claude Sonnet 5, Claude Opus 4.8, or GPT-5.2. The same write-up frames DeepMind's SynthID-Text as the original scheme Claude borrows from, which matters for editors and compliance teams: the detection tools tuned to SynthID's pattern may already be drifting as vendors iterate on sampling. The practical read is that "watermarked by default" should not be treated as a guarantee for any specific model version without re-running detection against that version.

Why vendors are splitting on watermark policy

Coverage on the same day frames Anthropic's watermark as designed to go undetected, while a separate note says DeepMind's watermark is marginally easier to detect and that applying it does not slow down text generation, and is said to outperform competitor schemes. That tension — invisibility versus reliable detectability — is the policy fault line now visible across the category: one vendor pushes invisible text marks as a default, while another is reported to have shipped a user-facing toggle in Gemini that lets people turn off the visible SynthID watermark on AI-generated images.

For a team standardising on one provider, the consequence is that provenance signals are no longer a uniform property of "AI text" or "AI images"; they are a per-product, per-modality decision that needs to be checked at the model level.

What editors and pipeline owners should verify next

Three concrete checks follow from the day's coverage. 2 explicitly; if you are on a different version, treat the watermark status as unknown. Second, audit any post-processing — translation, reformatting, summarisation — that touches Claude output, because the coverage notes the statistical signal can be erased and Tournament Sampling depends on the original token stream.

Third, revisit image-side disclosure assumptions for Gemini, since the visible SynthID watermark on AI-generated images is reported to have become opt-out via a toggle, meaning screenshots and exports no longer carry a guaranteed visible mark.

Tooling gaps the day's coverage implies

The reporting surfaces a handful of tool-shaped needs that readers in this space will look for next. One is a quick way to test whether a suspect text sample still carries the SynthID-style statistical signature after re-encoding, which lines up with a Unicode Encoder / Decoder workflow for normalising text before detection and a Reverse Words utility for stress-testing paraphrasing pipelines. Another is a side-by-side comparison helper for comparing visible versus invisible marks on AI images, where a Bold Text Generator or Vertical Text Generator can be used to mock up overlay variants for policy documentation, and a Punycode Converter helps when internationalised hostnames appear in provenance logs.

For teams that need to mark up internal reports, the Add Watermark to PDF capability and its plain-steps guide provide a reference pattern for layering visible provenance on top of invisibly signed AI text.