Claude's Watermark vs SynthID Text

Anthropic credits SynthID-Text, published by Google DeepMind in Nature in October 2024, as the source of the specific technique it uses. Both schemes descend from the same line of research: Scott Aaronson's 2022 proposal at OpenAI, the Kirchenbauer work in 2023, and the DeepMind paper that made it production-grade.

So the mechanism question has a short answer. They are the same idea. Everything that matters in practice is somewhere else.

The shared mechanism

Both systems tilt sampling toward tokens that a keyed pseudorandom function scores higher, chosen from the set of tokens the model already considers roughly equally good. The tilt is invisible to a reader. It is measurable by a party holding the key.

That single design decision produces every property the two schemes share.

What is common

• Nothing is added to the text, and there are no hidden characters in either scheme.
• Neither mark can be traced to a specific user, chat, or organization.
• Light editing does not remove either. A full rewrite removes both.
• Translation by the same model preserves the mark. Translation by a different model or a human removes it.
• Longer text scores more reliably. Short samples resolve to Unknown in both cases.

What differs

The key

Anthropic uses its own private key, so only Anthropic, and eventually Anthropic's detection API, can score text for the Claude watermark.

Google DeepMind uses its own private key for production Gemini output. The SynthID-Text repository on GitHub open-sources the code but not the production key. You can watermark your own text with your own key using that repository. You cannot use it to score Gemini output.

This is the detail that kills the most common shortcut. Running the open-source SynthID detector against Claude text, or against Gemini text, tells you nothing about the production mark, because it is not holding the key that produced it.

An independent test on August 23, 2026 illustrates both halves. Gemini 3.5 Flash tested positive for a SynthID-shaped signal. Sonnet 5 and Opus 4.8 did not carry one.

The surfaces marked

Claude marks text across the API, the Claude app, Claude Code, Cowork and Claude Tag, and attaches signed C2PA metadata to generated files.

SynthID covers more media types. Text in the Gemini app and API, an image mark that lives in the pixels rather than the metadata, plus audio and video variants across Google's generative model line.

Detector access

This is the sharpest difference, and it is the one that decides what anyone can actually do today.

Claude: no public detector. An API has been announced with no date.

SynthID Text: a detector portal exists, and it is waitlist-only for journalists, media professionals and researchers. There is no paste-and-check page for the public.

SynthID Image: verification is live in Google Search surfaces, is rolling out in Chrome, and is available as a limited-preview Cloud API. This is the only one of the four that an ordinary person can reach.

Model coverage rules

Claude marks models launched on or after August 2, 2026 at launch, with older models being retrofitted against an EU deadline of December 2, 2026. As of August 30, 2026 every shipped Claude model predates the cutoff.

Google states that essentially all Gemini text is watermarked, with no launch-date carve-out.

That difference explains why the two providers report so differently right now. Gemini text is probably marked. Claude text probably is not yet. Neither statement can be confirmed from a browser.

Which mark is on which text

• Text from Claude, the Anthropic API, Claude Code, Cowork, Tag, or Claude served through Bedrock, Vertex or Foundry: potentially Claude's mark, currently Unknown for every shipped model.
• Text from the Gemini app or Google's Gemini API: potentially SynthID Text, and likely present per Google's own statements and one independent test.
• Text from ChatGPT, the OpenAI API or Codex: no text watermark deployed as of August 30, 2026.
• Text from Copilot, Grok, Meta AI, Perplexity, Mistral or DeepSeek: no text watermark announced as of August 30, 2026.

The status matrix keeps this current, with the last-checked date on every row.

Where the two systems will diverge

Interoperability is the open question. If Anthropic and Google settle on a shared verifier interface, one check surface could cover both without either company exposing a key. If they do not, the industry lands where it is now: one detector per provider, each reachable only through that provider.

Enforcement surface is the other. Google is building SynthID checks into Search and Chrome, which puts a verification affordance in front of ordinary users. Anthropic has announced an API and no consumer surface at all. In terms of what actually reaches people, that difference matters more than anything about the mechanism.

And C2PA, which is neither

Content Credentials come up constantly in the same conversation and work on completely different principles. C2PA is a signed manifest stored inside a file. Any tool that understands the format can read it, including this one, and any tool that can write files can remove it.

That makes it the opposite of a statistical mark in almost every respect: readable by everyone, removable by anyone, and useless as a hidden signal. It is a label, not a watermark, and it is honest about being one.