How Anthropic and Google watermark AI text

You can hold a banknote up to the light and see the watermark. You cannot do that with a Claude or Gemini paragraph. The model does not append anything. It does not sprinkle zero-width characters or secret Unicode, and you are not billed for extra tokens. It still picks one word at a time. The change is which word wins when two options are both fine.

Both labs are reacting to the same pressure: the EU AI Act’s transparency rules, plus a July 2026 Code of Practice on marking AI-generated content that major providers signed. Anthropic said it cannot yet limit the mark by region, so new Claude models get it worldwide. Google has shipped SynthID on Gemini media for years, and SynthID-Text on Gemini chat since 2024.

Most Claude flagships you talk to today are still unmarked. Gemini text is usually watermarked.

The coin flip

Language models pick the next token, again and again. They do not print an essay in one shot. After “2 + 2 =” the next token should be “4”. After “The weather today was cold and…” it is a coin flip. Anthropic’s own example: “sugary” is a bad continuation. “Overcast” and “grey” are both fine. A reader will not care which one you get.

Watermarking takes those low-stakes flips. Instead of an arbitrary random number, the sampler uses a secret key plus a few words of context to decide which plausible word wins. Do that hundreds of times and a detector with the key can ask whether these choices lined up with the key more often than chance.

Anthropic’s method is a version of Google DeepMind’s SynthID-Text, published in Nature in 2024, in a line of work that goes back to Scott Aaronson’s 2022 proposal. Google describes the same idea as nudging probability scores. The open-source kit is a logits processor you can drop onto any Hugging Face model. Neither lab adds ink. They change the dice.

A useful picture, stolen from Anthropic: imagine playing Monopoly with the digits of pi instead of a die. The game still feels random. If you later see the whole sequence of moves, and you know pi, you can tell which rulebook was used.

Same email, different synonyms

Anthropic says internal tests showed no change in quality, creativity, or readability. DeepMind ran a live A/B on Gemini traffic and saw no statistically significant change in thumbs-up rates. Human raters, looking at watermarked and unwatermarked answers side by side, also saw no quality gap.

The two columns below are the same email with a handful of synonym swaps. Turn the highlights off. If you can pick the “AI” column every time, you are guessing.

Where the signal is thin

The watermark only lives in free choices, so it is uneven:

• Short text has too few flips, so confidence is low.
• Facts such as “Principia Mathematica” leave almost no room to wiggle.
• Code is mostly forced. Comments can carry a little signal. A working function usually cannot.
• Proofreading a human draft only marks the edits Claude actually made. Light copy-edits often will not register.
• Translations are fully chosen by the model, so they carry a mark.
• Rewrites that replace most words can wash it out. At that point it is fair to ask whether the text is still “AI-generated.”

Anthropic is explicit: a watermark cannot tell “Claude wrote this” from “Claude heavily edited this.” It also cannot tell you another lab wrote it. Different keys, sometimes different methods.

Which models are watermarked

Anthropic’s cutover date is 2 August 2026. Models launched on or after that day mark text at launch, on every surface Anthropic ships: the API, claude.ai, Claude Code, Cowork, Tag, and the same models on AWS, Google Cloud, and Microsoft Foundry. There is no opt-out.

The models people actually use on Chatio’s board (Opus 5, Sonnet 5, Fable 5, Haiku 4.5, the 4.x line) all launched before that date. They sit in the EU transition period. Anthropic says backfill is coming “over the coming months.” Independent checks in late August still found Sonnet 5 and Opus 4.8 looking unmarked.

Google started earlier, and the docs are messy. The SynthID product page still scopes text watermarking to the Gemini app and web experience. Images, video, and audio are marked across Google’s generative consumer products. You can ask Gemini whether an uploaded image, video, or audio file carries a Google SynthID. The public SynthID Detector portal is still media, not text.

Developer surfaces lagged in the docs. On the Gemini API forum, a Google reply first said API text was not watermarked, then corrected: API, Google AI Studio, and Antigravity text is SynthID-watermarked. A later public test of Gemini 3.5 Flash through AI Studio also came back marked. Take the correction as the current fact. The missing model-card sentence is a documentation gap.

Self-hosted Gemma weights do not include the watermark. SynthID-Text for Gemini lives in Google’s serving stack. If you run the open model yourself, you generate unmarked text unless you wire up the open-source logits processor.

OpenAI and others signed the same EU code. This piece is about the two labs that have published the mechanism in enough detail to show you the coin flip.

Files are a different kind of mark

When Claude writes a supported file (png, jpg, svg), it can attach a C2PA content credential: a signed note in metadata that says Claude processed the file. That is not a watermark. Strip the metadata, screenshot the image, or re-encode the file and the label can disappear. Google’s media SynthID is the opposite idea: the signal is in the pixels or samples, and it is built to survive crop, compression, and mild edits.

For assistants, the practical split is:

• Text you copy into email or docs carries a statistical watermark. It travels with the words and fades if you rewrite them.
• Images, audio, and video carry either metadata (Claude files) or SynthID in the media (Gemini, Imagen, Veo, Lyria, NotebookLM podcasts).

Neither mark changes who owns the output under the lab’s terms. Anthropic is clear on that. It is a provenance signal. It is not a copyright stamp, and it is not a user ID.

How you check

Anthropic says a watermark detection API is coming. It will return a likelihood that Claude was involved, not a courtroom proof. Google’s consumer check today is “upload the media to Gemini and ask.” There is still no simple public box where you paste a paragraph and get a SynthID-Text score.

Third-party “AI detectors” (Pangram and the rest) hunt style tells: the “this is not X, it is Y” cadence, the overused adverb. That is a classifier on vibes. A watermark check is a test against a key. Mixing the two is how people get false accusations.

Why this belongs on an assistant board

Chatio scores everyday help: did the model finish the job, follow the brief, sound like a person. Watermarking is not a quality axis. It is becoming part of the product. If you draft in Claude next month, after the backfill, that draft may still ping as Claude-touched after you paste it into a school portal or a CMS. If you draft in Gemini today, it likely already will. The coin flips are no longer private, and a statistical maybe is not a verdict about who wrote the sentence.

Primary sources: How Claude’s text watermark works, How Claude marks AI-generated content, SynthID, and SynthID Text.