Who uses watermark?

Public status checked on October 15, 2026. Date shows launch timing; — means no confirmed launch date.

Provider / framework Date Scope APIs Watermark algorithm
Google May 14, 2024 Gemini apps Text API: no watermark; detection: open-source tools.123 SynthID-Text: tournament sampling.4
Anthropic August 2026 Supported Claude models Generation: supported Claude APIs; detection: private preview.56 A version of SynthID-Text; production keys are private.5
OpenAI October 5, 2026 Selected models Generation: API opt-in; detection: restricted access.7 textGrain.7
vLLM September 24, 2026 Self-hosted inference Generation: Python / OpenAI-compatible APIs; detection: local utilities.89 Keyed Gumbel-max.8
SGLang — Self-hosted inference Generation: native / OpenAI-compatible APIs; detection: Python.10 Keyed Aaronson–Gumbel selection; optional dual-key mode.10

What is watermark?

Identify whether a passage was generated by my model.

Why do we use watermark?

  • Regulation: EU Code of Practice on Transparency of AI-Generated Content
  • Anti-LLMs: paper, homework, others

How to add watermark?

Naive way

It's a good day!

Authored by: CLAUDE

It's |CLAUDE| a |CLAUDE| good |CLAUDE| day |CLAUDE|!

Generative watermark (pseudo-random sampling)

Ordinary LLM generation with an unkeyed random seed generator, compared with keyed generative watermarking and watermark detection.
Adapted from Fig. 1 of Dathathri et al.; the ordinary-generation path now explicitly includes a random seed generator without a specified watermarking key. $x_{<t}$ is preceding text, $p_{\mathrm{LM}}(\cdot\mid x_{<t})$ is the next-token distribution, $r_t$ is the random seed, $x_t$ is the output token, and $x_1,\ldots,x_T$ is the text of length $T$ being scored.4 Editable Excalidraw.

Red-Green light (simple but not aligned with original LLM output)

  • Generate more green light words, less red light words
  • drawbacks：distortion
Red and green token assignments in unwatermarked and watermarked text, alongside token count, z-score and p-value.
Red/green token assignments and detector statistics in the supplied example. On narrow screens, swipe horizontally; click to view the original figure at full size.

Gumbel-max sampling

Gumbel-max sampling, followed by real Qwen3.5-4B int4 outputs with and without a watermark. Both texts are scanned with key kaiyu, synchronizing token highlights and cumulative evidence. English narration; approximately 2 minutes 16 seconds.

SynthID-Text (tournament sampling)

Most widely used methods (for better detection performance)

Published comparison of SynthID-Text and Gumbel sampling: true positive rate at one percent false positive rate versus token count.
Detection comparison in the supplied paper figure: TPR at a fixed 1% FPR versus text length.4
SynthID-Text random watermarking functions and a three-round tournament selecting an output token.
Tournament sampling: context and key generate random watermarking functions, which select among sampled candidates over successive rounds. Original Fig. 2, including its caption. On narrow screens, swipe horizontally.4

In-conclusion:

  • Higher entropy leads to better watermark
  • trade-offs between distortion and detectability

Challenges in Inference Systems

  1. Gumbel-max sampling overhead (negligible)
  2. Speculative Decode (Rejection sampling versus Gumbel-max sampling)

FAQs

Does watermarking affect LLM’s output?

No in theory.

No in Google’s 20 million Gemini thumbs-up and thumbs-down A/B tests.4

Anthropic reports no observed impact on content, creativity, or readability in its internal testing.5

References

  1. Google DeepMind, Watermarking AI-generated text and video with SynthID, May 14, 2024; Gemini app and web deployment announcement. ↩

  2. Google AI Developers Forum, Trey Nguyen, Does Gemini API text output carry SynthID watermarking?, August 5, 2026; API text scope in the support response. ↩

  3. Google DeepMind, SynthID Text, public implementation for self-hosted models; inspected October 15, 2026. ↩

  4. Sumanth Dathathri et al., Scalable watermarking for identifying large language model outputs, Nature 634, 818–823, October 23, 2024; Figs. 1–2, detection comparison and production evaluation. Paper figures are licensed CC BY 4.0; the overview was redrawn and extended with the ordinary random seed generator. ↩ ↩2 ↩3 ↩4 ↩5

  5. Anthropic, How Claude’s text watermark works, August 14, 2026; updated September 1, 2026 for detector access. ↩ ↩2 ↩3

  6. Anthropic, How Claude marks AI-generated content, model/surface coverage and partner rollout status inspected October 15, 2026. ↩

  7. OpenAI, Our approach to EU text provenance rules, October 5, 2026; API opt-in, EU rollout, textGrain and restricted detector access. ↩ ↩2

  8. Raphaël Rialland, Simon Veitner and Tomas Ruiz, Watermarking in vLLM, September 24, 2026. ↩ ↩2

  9. vLLM, Text watermarking, configuration, request surfaces and detection; inspected October 15, 2026. ↩

  10. SGLang, Text watermarking design, main snapshot 71a17db, inspected October 15, 2026; source-level evidence, not a release announcement. ↩ ↩2

  11. Scott Aaronson, Anthropic’s LLM watermarking, author commentary linked in the outline. ↩

  12. Zvi Mowshowitz, AI Text Watermarking Is Free And Good, community commentary linked in the outline. ↩