Which AI models watermark their text?
As of October 31, 2026: Google's Gemini app has watermarked text with SynthID since May 2024. Anthropic announced on August 14, 2026 that future Claude models carry a SynthID-based mark globally, with models released before August 2, 2026 to follow. OpenAI's textGrain, announced October 5, 2026, is on by default only for ChatGPT and Codex users in the EU, and opt-in for API customers worldwide. Open-weight models carry no provider mark; a deployer can add one with the open-source SynthID library. In every case the detector is held by the provider and access is limited to approved organisations.
The table
| Product | Watermark | Where it applies | User opt-out | Who can detect | Since |
|---|---|---|---|---|---|
| Gemini app and web | SynthID-Text | Everywhere | No | Google; eligible partners | May 2024 |
| Gemini API and Gemma | SynthID-Text available to developers | Developer's choice | Developer's choice | Whoever holds the key | October 2024 library |
| Claude (future models) | SynthID-Text variant | Everywhere | No | Anthropic; private-preview API for eligible organisations | Announced August 14, 2026 |
| Claude (models before August 2, 2026) | To be added | Everywhere, when added | No | Same | Over the months after the announcement |
| ChatGPT and Codex, EU users | textGrain | EU only | Not stated | OpenAI; approved researchers and expert organisations | Rolling out from October 5, 2026 |
| ChatGPT and Codex, outside the EU | None by default | n/a | n/a | n/a | n/a |
| OpenAI API | textGrain, opt-in, off by default | Worldwide, select models | Customer's choice | Same as above | October 5, 2026 |
| Open-weight models (Llama, Mistral, Gemma and others run by a deployer) | None unless the deployer adds one | Deployer's choice | Deployer's choice | Whoever holds the key | n/a |
Reading the table
Three facts explain most of the rows. First, all three production marks descend from SynthID-Text, so their strengths and limits are the same: the signal needs a few hundred words, fades with editing, and is gone after a full rewrite. Second, regional scope follows engineering, not policy preference: Anthropic went global because it could not scope by region; OpenAI chose an EU default and a worldwide opt-in because it could. Third, the detectors are all gated. There is no public tool that tells a reader whether a passage carries any of these marks, and each provider has said access will widen over time.
What is not on the table
Microsoft Copilot, Meta AI, Mistral's hosted products, xAI's Grok and the Chinese labs' consumer products are not listed because this directory had not verified a primary-source statement about text watermarking for them on the check date. Absence from the table means unverified, not unmarked. Image, audio and video marks are a separate matter: Google, OpenAI and Meta all mark media, and C2PA metadata is widely attached. If you can point to a primary source for a product not shown, the corrections address is open.
How this page is kept current
Each row is tied to an entry in the marks and provenance section, and each entry carries the date a person last checked the source. Changes are logged on the updates page and in the RSS feed. The same table is available to machines as part of the directory file described on the page for AI agents.
Questions people ask
Is ChatGPT output watermarked?
In the EU, yes for eligible ChatGPT and Codex output, rolling out from October 2026. Elsewhere, not by default. API customers anywhere can opt in.
Is Claude output watermarked?
Future Claude models carry a SynthID-based watermark worldwide, per Anthropic's August 14, 2026 announcement, with earlier models to follow.
Is Gemini output watermarked?
Yes. The Gemini app has applied SynthID to text since May 2024.
Do open-weight models like Llama watermark text?
Not by the model itself. Whoever deploys an open-weight model can add a watermark with a library such as SynthID-Text, but there is no provider mark.