Which AI models have confirmed text watermarking?

Compare confirmed text watermarking in Claude, Gemini, ChatGPT and Grok, including rollout dates, opt-in settings and important exceptions.

· 8 min read
Cover thumbnail with the text "AI text watermarking" and a waterdrop.

Claude, Gemini and ChatGPT have confirmed text watermarking, while supported Grok API models offer it as an opt-in setting. Coverage varies by model, region and product. The tables below show which outputs are covered and link directly to provider documentation.

Models with confirmed text watermarking

Last checked: November 11, 2026. Active means the provider documents watermarking as available for the listed output; Opt-in means it must be enabled, and Announced means availability has not yet been confirmed. Effective dates are included where documented; “Not disclosed” means an exact start date was not found in the linked source.

Anthropic: Claude

Supported models carry text watermarks across Claude products and cloud partners, worldwide. Anthropic’s current table lists the following models:

ModelStatusWatermarking methodEffective dateNotes
Claude Opus 5.5ActiveSynthID-Text variantNot disclosedRead more
Claude Opus 5ActiveSynthID-Text variantNot disclosedRead more
Claude Opus 4.8ActiveSynthID-Text variantNot disclosedRead more
Claude Opus 4.7ActiveSynthID-Text variantNot disclosedRead more
Claude Fable 5.2ActiveSynthID-Text variantNot disclosedRead more
Claude Fable 5.1ActiveSynthID-Text variantNot disclosedRead more
Claude Fable 5ActiveSynthID-Text variantNot disclosedRead more
Claude Mythos 5.2ActiveSynthID-Text variantNot disclosedRead more
Claude Mythos 5.1ActiveSynthID-Text variantNot disclosedRead more
Claude Mythos 5ActiveSynthID-Text variantNot disclosedRead more
Claude Sonnet 5.5ActiveSynthID-Text variantNot disclosedRead more
Claude Sonnet 5ActiveSynthID-Text variantNot disclosedRead more
Claude Sonnet 4.6ActiveSynthID-Text variantNot disclosedNon-default sampling settings can affect coverage. Read more
Claude Haiku 5.5ActiveSynthID-Text variantNot disclosedRead more

For Mythos 5, Opus 4.7 and Sonnet 4.6, Anthropic lists November 10 as the Amazon Bedrock rollout deadline. For Sonnet 4.6, omit non-default temperature, top_p and top_k settings if you rely on watermarking.

Google: Gemini

Model / productStatusWatermarking methodEffective dateNotes
Gemini app and webActiveSynthID TextMay 14, 2024API coverage is not established by this source. Read more

OpenAI: ChatGPT, Codex and API

Model / productStatusWatermarking methodEffective dateNotes
ChatGPTActivetextGrainNot disclosedEU text output. Read more
OpenAI API, select modelsOpt-intextGrainOctober 5, 2026Available worldwide; off by default. Read more
CodexAnnouncedtextGrainNot disclosedEU rollout announced October 5; completion not confirmed by the linked announcement. Read more

OpenAI’s announcement does not list individual API model IDs, so the table keeps its product-level scope.

xAI: Grok API

ModelStatusWatermarking methodEffective dateNotes
grok-4.7Opt-inSynthID TextNot disclosedAPI; off by default. Read more
grok-4.6Opt-inSynthID TextNot disclosedAPI; off by default. Read more
grok-4.5Opt-inSynthID TextNot disclosedAPI; off by default. Read more
grok-4.3Opt-inSynthID TextNot disclosedAPI; off by default. Read more
grok-4.20-reasoningOpt-inSynthID TextNot disclosedAPI; off by default. Read more
grok-4.20-non-reasoningOpt-inSynthID TextNot disclosedAPI; off by default. Read more
grok-4.20-multi-agentOpt-inSynthID TextNot disclosedAPI; off by default. Read more

Notable models without confirmed text watermarking

Unconfirmed does not mean unwatermarked. These entries distinguish missing confirmation from documented watermark support.

Provider / modelStatusNotes
Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5UnconfirmedAnthropic’s table leaves text-watermark columns blank. Read more
Meta: Llama / Meta AIUnconfirmedTextSeal is a research toolkit, not confirmation of production coverage. Read more
Mistral: Le Chat / APIUnconfirmedNo production text-watermark confirmation found in the provider documentation reviewed.
DeepSeek: hosted text outputUnconfirmedIts policy describes AI-generated labels, without establishing an embedded text watermark. Read more

For self-hosted open-weight models, watermarking also depends on how the operator deploys the model.

How do AI text watermarks work?

A text watermark is a signal the model embeds in its token-sampling probabilities; a detector is a third-party statistical guess with no model cooperation. They aren’t interchangeable.

Six concepts collapse into “watermarking”:

  • Statistical text watermarking biases token-sampling probabilities at generation time. Anthropic’s method is a SynthID-Text variant with no opt-out. Anthropic calls the impact negligible, a vendor claim, not independently benchmarked.
  • SynthID specifically is Google DeepMind’s peer-reviewed method, published in Nature, October 2024, using “Tournament Sampling” to modify token-sampling. A deployment of roughly 20 million Gemini responses showed no measurable quality loss.
  • Hidden or invisible Unicode characters (zero-width spaces, variation selectors) are an older, fragile technique that copy-paste or normalization strips without changing meaning. No first-party source confirms any model embeds this; some vendor blogs claim it, but cite no provider statement.
  • Content Credentials / C2PA are signed file-level metadata for images and video, not text, per Anthropic’s documentation, the clearest example of provenance that isn’t text watermarking.
  • AI-text detectors versus watermark detection: a detector guesses stylometrically; watermark detection matches a cryptographic pattern tied to a specific model. Turnitin’s false-positive rate is under 1% for papers over 20% AI-generated; a widely cited Stanford study found 61.3% average false positives across seven detectors on TOEFL essays by non-native English speakers. GPTZero self-reports 99.5% accuracy and 0.05% false positives; independent reviews report 80-90% accuracy and false positives up to about 10%. I’m flagging that inconsistency, not resolving it.
  • Platform labels and model-level provenance cover any signal tracing output to its system: a watermark, a C2PA credential, or a visible “AI-generated” badge. A label and an invisible watermark carry different guarantees; showing one doesn’t imply the other.

Can you check an AI text watermark?

Access depends on the provider:

  • Anthropic: a detection API is in private preview for eligible organizations; an interest form is available through its help page.
  • OpenAI: detector access requires approval for researchers and expert organizations. Apply through OpenAI’s announcement.
  • Google: public SynthID tools document checks for images, video and audio. Plain-text checking is not confirmed on the SynthID page.
  • xAI: detection API access is by application. For your own API responses, the x-watermark: true header reports that watermarking was applied. Read more.

General AI-writing detectors estimate whether text resembles AI output. That differs from checking a specific provider’s embedded watermark.

Is there a way to remove AI text watermarks?

No system reviewed here claims universal or guaranteed detection, and a missing watermark never proves human authorship.

Anthropic states plainly that detection is unreliable on short passages and sparse on factual or exact text. Light editing “probably won’t” remove the mark, but a full rewrite will; Claude’s translations remain watermarked too. A detected mark can’t tell “Claude wrote this” from “Claude edited or translated this,” and a missing mark doesn’t prove human authorship: it could mean an older model, heavy editing, short text, or an uncovered platform.

Independent testing is limited to open-source SynthID-Text implementations; no outside party can test the proprietary systems people actually use. Quality loss on prose didn’t exceed the effect of changing the sampling seed; code fared worse, with detectability near chance on both and correctness dropping on one model. Neither vendor robustness claims nor critics’ “easily removed” claims can be verified by outsiders, because no public tool can test deployed systems.

On paraphrasing and translation, one robustness study reported detection F1 dropping from 1.0 to 0.842 under DIPPER paraphrasing, with back-translation causing the worst degradation of any tested attack. A proposed mitigation recovered F1 to 0.923, showing defenses exist but aren’t standard yet. Older work showed a paraphrase attack estimating a watermark’s token pattern can defeat schemes like Unigram-Watermark and SIR while keeping output fluent, generalizing across schemes rather than one vendor.

On sampling and length: detection confidence rises with output length. OpenAI’s figures show about 80% detection at 200 tokens versus about 95% at 400, consistent with Anthropic’s short-passage caveat. No source quantified sampling temperature’s effect specifically. That’s an open question.

Update log

  • November 11, 2026: simplified the tables, refreshed provider documentation, updated Claude coverage, confirmed ChatGPT’s EU watermarking and added Grok API opt-in support.

Frequently Asked Questions

Does Claude watermark its text output?

Supported Claude models embed text watermarks across Claude products and cloud partners. Coverage depends on the model; Sonnet 4.6 has a limitation when non-default sampling settings are used.

Is ChatGPT's output watermarked?

OpenAI confirms textGrain watermarking for ChatGPT text in the EU. API customers worldwide can opt in; API watermarking is off by default.

Does text watermarking affect output quality?

Anthropic and Google say their text watermarks do not affect output quality. These are provider-reported claims.

Can I access an official AI text-watermark detector?

Access varies by provider. Anthropic offers a private preview for eligible organizations, while OpenAI and xAI require applications. Google's public SynthID tools document image, video and audio checks, rather than plain-text checks.

Summarize with AI

Track your AI search visibility

See where your brand appears in ChatGPT, Perplexity, and other AI search engines.

Start for free