Skip to main content
AI Watermark Removal

Technical guide

LLM Watermarking: Claude vs Gemini vs ChatGPT vs Llama

Six major providers, six different answers, all operating under the same new EU law. Anthropic marks Claude models launched on or after August 2, 2026, worldwide, and the first of them shipped on September 1, 2026. OpenAI built one reportedly around 99.9% effective and shelved it, then on October 5, 2026 announced a different one, textGrain, for ChatGPT and Codex text in the EU only, with an opt-in for API customers worldwide. Google says Gemini's works in the app and web experience, while a Google-affiliated developer answered a forum question by saying the API version doesn't exist, then reversed that on August 19, 2026 and said API text is watermarked after all. Grok, Llama, and DeepSeek aren't in the same conversation yet. Nobody is converging on a standard, and the law that was supposed to force one is explicitly technology-neutral.

By ignacio-sjPublished Revised Sources verified Research/proposal

Correction,

Four claims were wrong. The summary said Anthropic shipped a text watermark worldwide on August 2, 2026; August 2 is the launch-date threshold, and the first marked Claude models shipped on September 1. The Grok entry called xAI's Acceptable Use Policy clause the first confirmed sign that Grok output carries an embedded signal; the clause binds users and is not evidence of a mark, and the page now cites the Grok FAQ's visible image and video watermark. The OpenAI section said non-native English phrasing can resemble a watermark's statistical signature; the 2024 concern was that watermarking would stigmatise non-native speakers who use ChatGPT as a writing aid. The EU section attributed exemptions for source code, short sequences and machine-to-machine output to Article 50(2); the Article's own exemptions are standard editing assistance, outputs that do not substantially alter the input, and law-enforcement use, and the code and short-output carve-outs come from the Code of Practice.

All corrections

Key takeaways

  • No two major providers are in the same place: Claude has a confirmed watermark, Google states Gemini's covers the app and web (a claim the status table grades Contested) and the API is disputed, ChatGPT text is being watermarked in the EU only (textGrain, announced 2026-10-05 and rolling out over the coming weeks) and opt-in on the API, Grok and Llama have no confirmed deployment for ordinary chat text, and DeepSeek ships a visible label rather than an invisible one.
  • Most statistical watermarking, including Google's SynthID Text and the academic green-list scheme it descends from, biases token sampling during generation itself. That's mechanically different from metadata or hidden Unicode, both of which get attached after the text already exists.
  • OpenAI reportedly built a text watermark around 99.9% effective and held it back, per leaked internal documents citing false-positive risk, disproportionate impact on non-native English writers, and survey data suggesting roughly 30% of users would use ChatGPT less if it shipped. Its first shipped text watermark, textGrain, was announced on 2026-10-05 and is not a global default.
  • EU AI Act Article 50 creates real pressure to ship text watermarking, backed by fines up to €15M or 3% of global turnover. But the rule and its Code of Practice are explicitly technology-neutral, so provider responses will likely keep diverging rather than converge on one standard.

Provider map

Text watermark status by provider

Detector guide
How to read this map

Confirmed means an official or primary source documents text watermarking. Contested means official sources disagree with each other. Watchlist means regulation, research, or provider behavior makes the topic worth tracking closely.

Which LLM providers watermark their text?

Confirmed

A sourced position for each of the six providers, and why treating any two of them as equivalent is the most common mistake in this topic.

Provider status for ordinary generated text is not converging. It's diverging, and the gaps are documented in the providers' own words. The provider status tracker lists a sourced status for each one, and what AI watermarking is explains the mechanisms behind them.

  • Claude: confirmed, explained, and now partly checkable. Anthropic's support documentation says models launched on or after August 2, 2026 support machine-readable marking at launch, and its per-model table, re-read 2026-10-07, marks Fable 5.1, Mythos 5.1, Opus 5.5, Opus 5, Fable 5, Sonnet 5 and Opus 4.8. The remaining models released before that date have no Anthropic date; the AI Act gives them until December 2, 2026. Anthropic's 2026-09-25 email to Claude Platform customers dates Opus 5 from September 14, where its 2026-09-04 notice to administrators had said September 9, and dates Fable 5, Sonnet 5 and Opus 4.8 from September 30, 2026. Its August 14 explainer describes the mechanism as a version of Google DeepMind's SynthID-Text approach, keyed word selection with nothing added to the text. The detection API shipped on September 1 in private preview for regulators, media, fact-checkers, researchers and similar bodies, which made Claude the first provider on this page whose text watermark anyone outside the company could verify, and only if they qualify. OpenAI's textGrain detector, announced 2026-10-05, is likewise limited to approved researchers and expert organizations.
  • Gemini: stated for the app and web experience on Google's own SynthID page, unresolved for the API. A Google-affiliated developer-forum reply of August 5, 2026 said API text was not SynthID-watermarked; the same account said the opposite on August 19 and called its first answer incorrect. Google's documentation still covers neither claim.
  • ChatGPT: until 2026-10-05, nothing for ordinary text. On that date OpenAI announced textGrain, an invisible watermark rolling out over the coming weeks to ChatGPT and Codex text in the EU, across all plans, and not as a global default. API customers anywhere can opt in for select models; it stays off by default. The 2024 decision not to ship an earlier system is the most documented refusal story in the industry, and it gets its own section below.
  • Grok: xAI's Grok FAQ says generated images and videos carry a visible Grok watermark, and says nothing about text. Its Acceptable Use Policy prohibits stripping or circumventing "embedded provenance metadata or watermarks", but that clause is an obligation on users and does not show that xAI embeds any signal. xAI is not a C2PA member.
  • Llama: a full fetch of Meta's own Llama 3.1 and Llama 4 model cards found zero mentions of watermarking. Meta confirms watermarking elsewhere, Content Seal for Muse-generated images and a Llama Audio Watermark Detector for voice fraud, just nothing for chat text.
  • DeepSeek: a visible, human-readable label rather than an invisible statistical watermark, driven by China's Artificial Intelligence Generated and Synthetic Content Labeling Measures, effective September 2025. That's a different category of signal entirely, and easy to remove by deleting the label text.

Open-weight deployments add another layer of variance on top of all this. Whoever runs the actual model controls the sampler, so the same base model can be watermarked by one deployment and not another.

A 2024 paper by Christ, Gunn, Malkin, and Raykova proposes closing that gap by embedding the watermark in a model's parameters instead of relying on the deployer's sampling code. They argue it's provably resistant to token-substitution attacks unless output quality gets destroyed. That's a research proposal, not something any provider has shipped.

What does it mean for AI text to be watermarked?

Research/proposal

Here's the mechanical difference between the three things people call an AI text watermark, and why only one of them is immune to a text cleaner.

Most statistical LLM text watermarking, including SynthID Text and the academic green-list schemes that came before it, happens during generation. As the model builds a probability distribution over its next possible word, a watermarking layer nudges which word gets picked, using a pseudo-random signal derived from a secret key combined with the tokens that came just before.

The output still reads naturally, since the model keeps choosing among words it would plausibly pick anyway. It's just tilted toward some over others. A matching detector can later recover the statistical pattern from the sequence of choices.

Three unrelated things get called a text watermark, and they fail in completely different ways:

  • Statistical watermarking: baked into how the text was produced, so it can't be retrofitted onto text a differently configured model already generated.
  • Provenance metadata such as C2PA: a signed record attached to a file after the fact, routinely stripped by uploads, screenshots, and format conversion.
  • Hidden Unicode characters: inserted into finished text, and removable by any tool that normalizes characters.

Hidden characters usually survive a paste into a plain text box too, which is why character cleaners exist. Only the first kind survives a cleaner, because there is no character to strip.

What happened to OpenAI's shelved text watermark?

Reported

Why the most effective text watermark anyone has reportedly built never shipped, in the four reasons OpenAI itself gave for holding it back.

Per Wall Street Journal reporting on leaked internal documents, relayed by Tom's Hardware, OpenAI built a text-watermarking method roughly 99.9% effective at flagging ChatGPT-written text. Per that 2024 reporting, the company had weighed releasing it for about a year.

It never went out the door. The internal case against release came down to four things:

  • Real risk of false positives against innocent human writers.
  • A concern that watermarking would stigmatise non-native English speakers who use ChatGPT as a writing aid.
  • Straightforward circumvention once the method became public.
  • Survey data suggesting close to 30% of users would use ChatGPT less if the feature shipped.

Until 2026-10-05, nothing superseded this: OpenAI described text provenance as a future goal while shipping watermarking for images (SynthID plus C2PA, since May 2026) and for audio (since July 2026).

On 2026-10-05 OpenAI announced textGrain, a different scheme, and released it narrowly: an API opt-in that stays off by default, with ChatGPT and Codex text in the EU rolling out over the coming weeks, and the detector limited to approved researchers. OpenAI publishes figures for how synonym swaps weaken it, and says substantial paraphrasing or translation can make the mark undetectable.

Does the EU AI Act require a specific watermark?

Official announcement

Article 50 sets a duty with three exemptions and names no scheme. The code and short-output carve-outs come from the Code of Practice, not the Article.

EU AI Act Article 50(2) requires providers of systems generating synthetic text to make outputs machine-readable and detectable as artificially generated, as far as technically feasible. That rule took effect August 2, 2026, backed by fines up to €15M or 3% of global annual turnover.

Roughly 190 organizations, including Anthropic, Google, Meta, Microsoft, and OpenAI, had signed the EU's Code of Practice on transparency by the end of July 2026.

But the rule and its Code of Practice are both explicitly technology-neutral. Neither names a required scheme, and Article 50(2) itself exempts three things:

  • AI systems performing an assistive function for standard editing
  • Outputs that do not substantially alter the input data or its meaning
  • Uses authorised by law to detect, prevent, investigate or prosecute criminal offences

The Code of Practice adds its own carve-outs. As OpenAI summarises it, the Code does not require marks on outputs under 200 tokens or on code snippets.

So provider responses are likely to keep diverging rather than converge on one shared standard. And because the Code does not require marks on code snippets, the code-safety question about token-nudging watermarks stays as open as it was before.

FAQ

Will every LLM use the same watermark?

No. Claude has a confirmed mechanism, Google states Gemini's covers the app and web while the API is disputed, ChatGPT text carries OpenAI's textGrain watermark in the EU only (announced 2026-10-05, rolling out over the coming weeks) and on the API only when a customer opts in, Grok and Llama have nothing confirmed for ordinary chat text, and DeepSeek uses a removable visible label instead of a statistical watermark. Open-weight deployments add still more variance, since the operator running the model controls whether any watermarking gets applied at all.

Is Gemini's text watermark confirmed or not?

Stated by Google for the app, unresolved for the API. Google's own SynthID page states that Gemini app and web text is watermarked, though nobody outside Google can check it. For the API, a Google-affiliated forum account denied it on August 5, 2026 and reversed itself on August 19, saying API text "IS actually SynthID-watermarked" and extending that to AI Studio and Antigravity. Neither post is documentation, no Google page covers the API either way, and no published tool accepts text for checking, so this is an open question rather than a settled answer.

Why would OpenAI build a working watermark and then not release it?

Per Wall Street Journal reporting on leaked internal OpenAI documents, the company's own case cited false-positive risk, disproportionate impact on non-native English writers, easy circumvention once the method became public, and survey data suggesting roughly 30% of users would use ChatGPT less if it shipped. That's the most detailed public rationale any major provider has given for withholding a text watermark. On 2026-10-05 OpenAI announced a different one, textGrain, limited to the EU for ChatGPT and Codex and opt-in on the API.

Does a provider's watermarking policy apply to self-hosted open-weight models?

No. Whoever runs the actual inference controls the token sampler, so watermarking on an open-weight model depends entirely on how that specific operator configured their deployment, regardless of what the model's original publisher does on its own hosted service.

Next steps

  • Read Google's SynthID page and the developer-forum reply back to back. It takes five minutes and it's the fastest way to see how genuinely unresolved the Gemini API question is. Is Gemini watermarked?
  • Get the full OpenAI picture, including the image and audio provenance it did ship and the C2PA stripping bug users found in ChatGPT itself. Is ChatGPT watermarked?
  • Read how OpenAI's textGrain text watermark works, who gets it, and what OpenAI's own numbers say it survives. OpenAI's textGrain watermark
  • Go one level deeper on the mechanism and see exactly where in the sampling pipeline the nudge happens. Claude's watermark against known schemes
  • If you're trying to comply rather than just understand, start with what Article 50 actually asks for and what it exempts. EU AI Act and AI watermarking

Sources and citation status