What SynthID-Text is and how statistical text watermarking works

Google shipped the first text watermark and documented it. Understanding SynthID explains how Anthropic's new Claude mark most likely behaves too.

Which providers watermark · 2 min read

Google's SynthID-Text is the best-documented text watermark in existence: published in Nature in October 2024 and running in Gemini. Anthropic began marking Claude's text in August 2026 but has not yet published its mechanism, so SynthID remains the clearest window into how these systems behave.

How it works

Text generation is a sequence of choices. At each step the model has a probability distribution over possible next tokens and samples one. Usually several tokens are nearly equally good.

SynthID exploits that slack. A pseudorandom function seeded by a secret key assigns scores to candidate tokens, and sampling is nudged toward higher-scoring ones. The nudge is small enough that quality does not visibly suffer, but across hundreds of tokens the accumulated bias forms a statistical pattern.

A detector holding the same key re-computes the expected scores and measures whether the text scores higher than chance. Enough excess, and it reports the text as likely watermarked.

Google calls this "tournament sampling" — candidates compete in bracket rounds scored by the key.

What it can and cannot prove

Can: give a confidence score that a reasonably long passage came from a watermarked model.

Cannot:

• Work on short text. A tweet has too few token choices to accumulate signal.
• Survive heavy paraphrasing. Rewriting replaces the very choices that carried the signal.
• Survive translation.
• Tell you who generated it, or when, or with what prompt.
• Work at all without the key. This is not a public checksum — only the key holder can detect it.

That last point is the one most people miss. A watermark is not a visible mark on the text. It is a private statistical property.

Why it matters for the rest of the field

SynthID is the existence proof that text watermarking is practical at scale, and the honest illustration of the limits: the same paper that demonstrates it works also documents that editing degrades it.

That trade-off is why OpenAI held its own system back. It is also why Anthropic's documentation for Claude's mark lists heavy editing, paraphrasing and translation as things it may not survive — the same weakness, because it is inherent to the approach rather than to one implementation.

What this means for you

Gemini output: a SynthID signal may be present, and only Google can read it.

Claude output from models launched on or after 2 August 2026: a mark is present. Anthropic has not published the detection mechanism, so nobody outside Anthropic can currently check for it.

GPT output: no deployed watermark as of this writing.

In all three cases, what a scan can count exactly is the character layer: zero-width characters, exotic spacing, typographic punctuation, and stylistic markers.

Scan your text to see exactly which of those you have.

Related: does Anthropic watermark Claude's text? · can you remove it? · does ChatGPT watermark text?