SynthID text watermark and how to remove it

It is a statistical signature in which words a model picks, steered by a secret key, with no character added anywhere in the text.

Google's developer docs call SynthID Text "a logits processor": it tips the pick after the model has ranked its likely next words. The Nature paper of 23 October 2024 explains Google's method, called tournament sampling; Anthropic says only that Claude uses a version of the approach, without naming the steps. As Google publishes them:

1. The secret key and the last few words seed a hidden scoring rule.
2. The model draws several candidate next words.
3. They meet in pairs over knockout rounds; the higher score goes through.
4. The last word standing is written.

A detector holding the same key scores every word of a passage and averages the result. One word proves nothing; hundreds add up. Google advises keeping each configuration private, since otherwise a watermark "may be trivially replicable by others."

Google DeepMind and Hugging Face released the code openly on 23 October 2024 under the Apache-2.0 license, for use with keys you create yourself. Since the mark sits in word choice, a tool that deletes invisible characters cannot reach it (our reading of Anthropic's statement that nothing is added to Claude's text).

Sources: Dathathri et al., Nature 2024 · Google developer docs · Hugging Face, 23 Oct 2024 · google-deepmind/synthid-text · Anthropic explainer

Google's Gemini app and web experience since May 2024, and Claude, whose watermark Anthropic calls a version of the SynthID-Text approach.

Google announced text marking for the Gemini app and web on 14 May 2024. The Nature paper says SynthID-Text has marked both Gemini and Gemini Advanced, and reports a live check across nearly 20 million responses. Anthropic applies its version at the model level, so it is present whichever Claude product the text comes from, wherever Claude is offered. It also covers translations Claude makes, since the model picks every word of the translated text.

OpenAI marks ChatGPT text with a different system under its own name; ChatGPT's separate text watermark is covered on our sister site.

Sources: Google DeepMind, 14 May 2024 · Dathathri et al., Nature 2024 · Claude help center

Only a party holding the key, and on the day we checked our sources, 19 Oct 2026, no public tool read SynthID in text.

For Claude, Anthropic runs a detection API in private preview for eligible organizations such as regulators, media, fact-checkers, researchers and schools, and its key answers one narrow question: how likely the text was "partly written by Claude."

For Gemini, Google opened its SynthID Detector to everyone on 7 October 2026, but its page lists three supported content types: images, video and audio. Text is not one of them.

Anyone can run the open-source detector, but only on text marked with a configuration they hold, never on Google's or Anthropic's production marks. Style-based checkers guess from phrasing and hold no key at all; the Claude AI detector page explains what their scores mean for Claude text.

So a site that claims to spot SynthID in pasted text is reading neither company's key, and no public tool can confirm a SynthID result in text.

Sources: Anthropic explainer · Google, 7 Oct 2026 · SynthID Detector · Google developer docs

Yes, roughly in step with how much wording changes: Google says the mark holds up under "a few words and mild paraphrasing," while its confidence "can be greatly reduced" after a thorough rewrite or a translation.

The table beside this comes from Liang et al., "Watermark under Fire" (arXiv 2411.13425, version 4, September 2025), a preprint that tested open-source SynthID-Text on Qwen2.5-14B-Instruct with its Paper Conclusion dataset at a 1% false-positive rate. The paraphrases came from the DIPPER paraphraser at a weaker setting (DP-20) and a stronger one (DP-40); the translation went into French and back.

Read it as a slope, not a forecast of what happens to a live Gemini or Claude mark. Anthropic's help center lists the same causes for a missing mark, text that was "heavily edited, paraphrased, translated, or mixed into other writing," and its explainer says light editing probably leaves part of the pattern behind.

Sources: Liang et al., arXiv 2411.13425v4 · Google DeepMind · Claude help center · Anthropic explainer

Change the wording, not the characters: a rewrite that replaces the word choices is the one change both Google and Anthropic point to.

Changing the wording is the job of the rewrite tab. Our rewriting engine rebuilds your passage in new wording, then prints a receipt with five lines: Words in, Words out, Hidden characters removed, Original 3-word runs kept (as a percentage) and Words left. The kept figure is the one to watch: the lower it is, the more of the phrasing has changed. It measures your wording, not the mark itself.

Neither company gives a word count to aim for. Google speaks of text "thoroughly rewritten", Anthropic of every word replaced. The Claude side of this, model by model, is on the Claude watermark remover page.

If you wrote the text yourself and only asked Claude to proofread it, any mark can only sit in the few corrections the model made. Stray invisible characters from copying between apps are a separate matter; the free cleaning tab strips them.

Sources: Anthropic explainer · Google developer docs · Liang et al., arXiv 2411.13425v4

Less well, because with few real word choices there is little to read: Anthropic says small samples give "less information to go on," and Google says factual prompts leave fewer chances to adjust the wording.

The Nature paper states the general rule: the method "is less effective when there is less entropy," meaning when the model has only one or two sensible next words. Anthropic lists the same weak spots for Claude: factual passages carry a sparser mark, code generally less, and a proofread text holds it only in "the handful of corrections."

Length works the other way: a long, freely written essay carries far more of the pattern than a two-line reply. Neither company publishes a minimum length.

Sources: Dathathri et al., Nature 2024 · Anthropic explainer · Google DeepMind

No. These are three different things: a word-choice pattern in text, a separate watermark inside the pixels, frames or sound of media, and a signed note in the metadata of files Claude produces.

Google's public SynthID Detector reads the media watermark, which Google says sits in over 180 billion images and videos. This tool works on text only.

Content Credentials are something else again. Anthropic calls that label "very different from a watermark. Nothing in the file changes." Claude's developer docs add that re-encoding, a format change, a screenshot or a metadata-stripping tool each removes the credential.

Sources: Google, 7 Oct 2026 · SynthID Detector · Anthropic explainer · Claude developer docs

Four free rewrites of up to 250 words after an email sign-in, then plans from $12 a month on yearly billing; stripping hidden characters is always free.

Monthly plans are Basic $18, Pro $27 and Ultra $48; paid yearly they come to $144, $216 and $432. Every rewrite ships with its receipt. See the prices