Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 30 of the 30 references it cites were found in a registry or an index; 11 of the 26 sources its claims cite were checked for support.

Scalable watermarking for identifying large language model outputs

Open access

Abstract

Large language models (LLMs) have enabled the generation of high-quality synthetic text, often indistinguishable from human-written content, at a scale that can markedly affect the nature of the information ecosystem 1–3 . Watermarking can help identify synthetic text and limit accidental or deliberate misuse 4 , but has not been adopted in production systems owing to stringent quality, detectability and computational efficiency requirements. Here we describe SynthID-Text, a production-ready text watermarking scheme that preserves text quality and enables high detection accuracy, with minimal latency overhead. SynthID-Text does not affect LLM training and modifies only the sampling procedure; watermark detection is computationally efficient, without using the underlying LLM. To enable watermarking at scale, we develop an algorithm integrating watermarking with speculative sampling, an efficiency technique frequently used in production systems 5 . Evaluations across multiple LLMs empirically show that SynthID-Text provides improved detectability over comparable methods, and standard benchmarks and human side-by-side ratings indicate no change in LLM capabilities. To demonstrate the feasibility of watermarking in large-scale-production systems, we conducted a live experiment that assessed feedback from nearly 20 million Gemini 6 responses, again confirming the preservation of text quality. We hope that the availability of SynthID-Text 7 will facilitate further development of watermarking and responsible use of LLM systems.

Overview

CoChat Summary

TL;DR

SynthID-Text is a production-ready generative text watermarking scheme that modifies only the LLM sampling procedure, using a novel Tournament sampling algorithm to preserve text quality while improving detectability over comparable methods. It was deployed in Gemini and Gemini Advanced and evaluated on nearly 20 million live responses, with negligible latency overhead.

Key findings

  • Non-distortionary SynthID-Text preserved quality in a live Gemini experiment over approximately 20 million watermarked and unwatermarked responses: thumbs-up rate differed by 0.01% (watermarked higher) and thumbs-down rate by 0.02% (watermarked lower), both statistically insignificant.
  • Non-distortionary SynthID-Text provided better detectability than Gumbel sampling for the same text length, measured as true-positive rate at 1% false-positive rate on Gemma 7B-IT with temperature 0.7 and ELI5 prompts; distortionary SynthID-Text offered a more favourable detectability/quality trade-off than Soft Red List.
  • Computational impact was minimal: Gemma 7B-IT on 4 v5e TPUs generated text at 15.527 ms per token, increasing to 15.615 ms per token with 30-layer Tournament sampling, a latency increase of only 0.57% (Gumbel 0.26%, Soft Red List 0.28%).
  • A controlled human preference test on 3,000 ELI5 questions comparing watermarked versus unwatermarked Gemma 7B-IT responses found no significant difference across grammaticality/coherence, relevance, correctness, helpfulness, and overall quality.

Methods

  • SynthID-Text uses generative watermarking: a random seed generator, a sampling algorithm, and a scoring function. It uses Tournament sampling, with a sliding-window random seed generator hashing the last H = 4 tokens plus a watermarking key, m = 30 tournament layers, Bernoulli(0.5) g-values, and a Bayesian scoring function; detection requires only tokenized text, the key, and the seed generator, not the LLM.
  • Evaluations used instruction-tuned Gemma 2B and Gemma 7B, and Mistral 7B-IT, with top-k = 100 and temperatures 0.5, 0.7, and 1.0, prompted by ELI5 questions. Non-distortionary experiments used 10,000 disjoint prompts each for test and development; distortionary experiments used 1,500 ELI5 prompts for the test set. Detectability was reported as true-positive rate at a fixed false-positive rate.
  • The live production experiment routed a random fraction of Gemini queries to a watermarked model and an equivalent number to an unwatermarked counterpart, analysing approximately 20 million responses via thumbs-up and thumbs-down feedback. Additional automatic evaluations checked perplexity and standard benchmarks.

Limitations

  • The paper states that generative watermarks such as SynthID-Text do not offer a complete solution to AI text detection; they require coordination between actors running LLM services, are difficult to enforce on decentralized open-source models, and are vulnerable to stealing, spoofing, and scrubbing attacks. They are also weakened by edits such as LLM paraphrasing.
  • The paper states that no text detection method is foolproof and that watermarking is complementary to other approaches; detecting text from actors who do not implement watermarking requires other methods such as post hoc detection.
  • Evident from the text: detection depends on text length and LLM distribution entropy, performing better with more entropy and less effectively in low-entropy settings; the non-distortionary configuration has some reduction to inter-response diversity, and the distortionary configuration trades text quality for higher detectability.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41586-024-08025-4

Cites

30 of 30 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Overview of LLM text generation and generative watermarking. Top: LLM text generation typically involves generating text from left to right by repeatedly sampling from the LLM distribution. Bottom: a generative watermarking scheme typically consists of the three components, in the blue boxes: random seed generator, sampling algorithm and scoring function. These can be used to provide a text generation method and a watermark detection method. In the SynthID-Text generative watermarking scheme, we use the Tournament sampling algorithm.
Fig. 1 Overview of LLM text generation and generative watermarking. Top: LLM text generation typically involves generating text from left to right by repeatedly sampling from the LLM distribution. Bottom: a generative watermarking scheme typically consists of the three components, in the blue boxes: random seed generator, sampling algorithm and scoring function. These can be used to provide a text generation method and a watermark detection method. In the SynthID-Text generative watermarking scheme, we use the Tournament sampling algorithm.CC BY · PubMed Central
Fig. 2 SynthID-Text’s Tournament-based watermarking. Top: to generate a new token xt, we first score each token in the vocabulary using multiple (in this case, m = 3) random watermarking functions g1, …, gm. These assign random values using a random seed, which is generated based on both the recent context and a watermarking key. Bottom: then, we choose the next token using a tournament process. First, we sample 2m = 8 (possibly non-unique) tokens from pLM(⋅∣x<t). These are split into pairs of competing tokens; in each pair, the highest scoring one (based on g1) is chosen, breaking ties randomly. The resulting tokens compete in the next layer, where winners are chosen based on g2, until in the last tournament layer the final winner is selected based on gm: this becomes the next generated token xt.
Fig. 2 SynthID-Text’s Tournament-based watermarking. Top: to generate a new token xt, we first score each token in the vocabulary using multiple (in this case, m = 3) random watermarking functions g1, …, gm. These assign random values using a random seed, which is generated based on both the recent context and a watermarking key. Bottom: then, we choose the next token using a tournament process. First, we sample 2m = 8 (possibly non-unique) tokens from pLM(⋅∣x<t). These are split into pairs of competing tokens; in each pair, the highest scoring one (based on g1) is chosen, breaking ties randomly. The resulting tokens compete in the next layer, where winners are chosen based on g2, until in the last tournament layer the final winner is selected based on gm: this becomes the next generated token xt.CC BY · PubMed Central
Fig. 3 Detection performance of SynthID-Text. a, For non-distortionary watermarking, we compare watermark detectability as a function of text length (number of tokens), for non-distortionary SynthID-Text and the Gumbel sampling22,24 watermark. Watermark detectability is measured using the true-positive rate (TPR) when the false-positive rate (FPR) is set to 1%. Responses are generated from Gemma 7B-IT with temperature = 0.7 using prompts from the ELI5 dataset; other models and temperatures are provided in Extended Data Fig. 1. b, The fraction of watermarked and unwatermarked texts that the selective prediction mechanism (Supplementary Information section C.8) abstains on to achieve a true-positive rate of 95% and a false-positive rate of 1%. The mechanism makes predictions only when the scoring function is confident, and is thus able to maintain a low error rate when making predictions, at the cost of abstaining on some of the data. The model set-up and the prompts are same as that in a; other temperatures are in Extended Data Fig. 3. c, For distortionary watermarking, we compare detectability of watermarks that allow text quality to be traded off against detectability by adjusting a strength hyperparameter. Texts are of length 200 tokens generated by Gemma 7B-IT with temperature = 0.7; other text lengths and temperatures are provided in Extended Data Fig. 2. Compared with the distortionary Soft Red List watermark23, distortionary SynthID-Text offers a more favourable trade-off, with substantially higher detection rates for the same effect on text quality as measured by log(perplexity). The arrows indicates the direction for performance improvement. The dashed lines correspond to a bootstrap estimate (500 resamples) of the mean of the metric on the y axis and the shaded regions correspond to the 90% confidence interval on the mean estimate.
Fig. 3 Detection performance of SynthID-Text. a, For non-distortionary watermarking, we compare watermark detectability as a function of text length (number of tokens), for non-distortionary SynthID-Text and the Gumbel sampling22,24 watermark. Watermark detectability is measured using the true-positive rate (TPR) when the false-positive rate (FPR) is set to 1%. Responses are generated from Gemma 7B-IT with temperature = 0.7 using prompts from the ELI5 dataset; other models and temperatures are provided in Extended Data Fig. 1. b, The fraction of watermarked and unwatermarked texts that the selective prediction mechanism (Supplementary Information section C.8) abstains on to achieve a true-positive rate of 95% and a false-positive rate of 1%. The mechanism makes predictions only when the scoring function is confident, and is thus able to maintain a low error rate when making predictions, at the cost of abstaining on some of the data. The model set-up and the prompts are same as that in a; other temperatures are in Extended Data Fig. 3. c, For distortionary watermarking, we compare detectability of watermarks that allow text quality to be traded off against detectability by adjusting a strength hyperparameter. Texts are of length 200 tokens generated by Gemma 7B-IT with temperature = 0.7; other text lengths and temperatures are provided in Extended Data Fig. 2. Compared with the distortionary Soft Red List watermark23, distortionary SynthID-Text offers a more favourable trade-off, with substantially higher detection rates for the same effect on text quality as measured by log(perplexity). The arrows indicates the direction for performance improvement. The dashed lines correspond to a bootstrap estimate (500 resamples) of the mean of the metric on the y axis and the shaded regions correspond to the 90% confidence interval on the mean estimate.CC BY · PubMed Central
Extended Data Fig. 1 Detectability comparison of non-distortionary SynthID-Text vs Gumbel sampling, for additional model and temperature combinations. We show a range of temperatures from 0.5–1.0, which is the range most often used in practice. SynthID-Text generally provides improved detection over Gumbel sampling, with the improvements being greatest in lower entropy settings (lower temperature and larger models). The dashed lines correspond to a bootstrap estimate (500 resamples) of the mean true positive rate (TPR) at a false positive rate of 1%, and the shaded regions correspond to the 90% confidence interval on the mean estimate.
Extended Data Fig. 1 Detectability comparison of non-distortionary SynthID-Text vs Gumbel sampling, for additional model and temperature combinations. We show a range of temperatures from 0.5–1.0, which is the range most often used in practice. SynthID-Text generally provides improved detection over Gumbel sampling, with the improvements being greatest in lower entropy settings (lower temperature and larger models). The dashed lines correspond to a bootstrap estimate (500 resamples) of the mean true positive rate (TPR) at a false positive rate of 1%, and the shaded regions correspond to the 90% confidence interval on the mean estimate.CC BY · PubMed Central
Extended Data Fig. 2 Detectability and text quality comparison of distortionary SynthID-Text and Soft Red List, for additional temperatures and text lengths. Detectability is measured by true positive rate at a fixed false positive rate of 1% (TPR@FPR=1%), and text quality is measured by log perplexity. Texts are generated by Gemma 7B-IT. See Supplementary Information section D for details. Dashed lines correspond to a bootstrap estimate (500 resamples) of the mean TPR@FPR=1%, and the shaded regions correspond to the 90% confidence interval on the mean estimate.
Extended Data Fig. 2 Detectability and text quality comparison of distortionary SynthID-Text and Soft Red List, for additional temperatures and text lengths. Detectability is measured by true positive rate at a fixed false positive rate of 1% (TPR@FPR=1%), and text quality is measured by log perplexity. Texts are generated by Gemma 7B-IT. See Supplementary Information section D for details. Dashed lines correspond to a bootstrap estimate (500 resamples) of the mean TPR@FPR=1%, and the shaded regions correspond to the 90% confidence interval on the mean estimate.CC BY · PubMed Central
Extended Data Fig. 3 Abstention rates for non-distortionary SynthID-Text, for additional temperatures. The fraction of watermarked and unwatermarked texts abstained on by the selective prediction mechanism described in Supplementary Information section C.8, for Gemma 7B-IT. The mechanism abstains when the scoring function is uncertain, and is configured here to obtain a true positive rate of 95% and a false positive rate of 1% on the remaining texts. The abstention rate is lower for longer texts and higher temperatures. Dashed lines correspond to a bootstrap estimate (500 resamples) of the mean abstention rate on watermarked and unwatermarked texts, and the shaded regions correspond to the 90% confidence interval on the mean estimate.
Extended Data Fig. 3 Abstention rates for non-distortionary SynthID-Text, for additional temperatures. The fraction of watermarked and unwatermarked texts abstained on by the selective prediction mechanism described in Supplementary Information section C.8, for Gemma 7B-IT. The mechanism abstains when the scoring function is uncertain, and is configured here to obtain a true positive rate of 95% and a false positive rate of 1% on the remaining texts. The abstention rate is lower for longer texts and higher temperatures. Dashed lines correspond to a bootstrap estimate (500 resamples) of the mean abstention rate on watermarked and unwatermarked texts, and the shaded regions correspond to the 90% confidence interval on the mean estimate.CC BY · PubMed Central
Extended Data Fig. 4 Comparing the tradeoff between diversity and detectability for non-distortionary SynthID-Text and Gumbel sampling. Diversity is indicated by low Self-BLEU and detectability is measured by true positive rate when false positive rate is fixed at 1% (TPR@FPR=1%). Texts are length 400 tokens, generated from Gemma 7B-IT at three different model temperatures. We observe that SynthID-Text provides a more favorable trade-off across temperatures. Dashed lines correspond to a bootstrap estimate (500 resamples) of the mean TPR@FPR=1%, and shaded regions correspond to the 90% confidence interval on the mean estimate. See Supplementary Information section C.3 for details.
Extended Data Fig. 4 Comparing the tradeoff between diversity and detectability for non-distortionary SynthID-Text and Gumbel sampling. Diversity is indicated by low Self-BLEU and detectability is measured by true positive rate when false positive rate is fixed at 1% (TPR@FPR=1%). Texts are length 400 tokens, generated from Gemma 7B-IT at three different model temperatures. We observe that SynthID-Text provides a more favorable trade-off across temperatures. Dashed lines correspond to a bootstrap estimate (500 resamples) of the mean TPR@FPR=1%, and shaded regions correspond to the 90% confidence interval on the mean estimate. See Supplementary Information section C.3 for details.CC BY · PubMed Central

Verification

Claims the paper makes

11 Valid · 4 Not valid · 12 Not checked · 60 Own results and interpretations

45 more claims

Create a free account to see all 87 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 27
  • Probably not a substantial claim. 16

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 24 references

Research works printed but not linked: 7

  1. Taori, R. & Hashimoto, T. Data feedback loops: model-driven amplification of dataset biases. In Proc. 40th International Conference on Machine Learning 33883–33920 (JMLR, 2023).
  2. Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D. & Finn, C. DetectGPT: zero-shot machine-generated text detection using probability curvature. In Proc. 40th International Conference on Machine Learning 24950–24962 (JMLR, 2023).
  3. Kuditipudi, R., Thickstun, J., Hashimoto, T. & Liang, P. Robust distortion-free watermarks for language models. Trans. Mach. Learn. Res. https://openreview.net/pdf?id=FpaCL1MO2C (2024).
  4. Christ, M., Gunn, S. & Zamir, O. Undetectable watermarks for language models. In Proc. Thirty Seventh Conference on Learning Theory 1125–1139 (PMLR, 2024).
  5. Jovanović, N., Staab, R. & Vechev, M. Watermark stealing in large language models. In Proc. 41st International Conference on Machine Learning 22570–22593 (PMLR, 2024).
  6. Zhang, H. et al. Watermarks in the sand: impossibility of strong watermarking for language models. In Proc. 41st International Conference on Machine Learning 58851–58880 (PMLR, 2024).
  7. Holtzman, A., Buys, J., Du, L., Forbes, M. & Choi, Y. The curious case of neural text degeneration. In Proc. Eighth International Conference on Learning Representations (ICLR, 2020).

Cited by

See all 132

Metadata

Authors
Dathathri, S., See, A., Ghaisas, S., Huang, P.-S., McAdam, R., Welbl, J., Bachani, V., Kaskasoli, A., Stanforth, R., Matejovicova, T., Hayes, J., Vyas, N., Merey, M. A., Brown-Cohen, J., Bunel, R., Balle, B., Cemgil, T., Ahmed, Z., Stacpoole, K., Shumailov, I., Baetu, C., Gowal, S., Hassabis, D., & Kohli, P.
Affiliations
  • Google DeepMind (United Kingdom) (GB)
  • Google (United States) (US)
Corresponding author
  • Dathathri, Sumanth
  • Kohli, Pushmeet
Venue
Nature
Published
2024-10-23
Volume
634
Issue
8035
Pages
818-823
ISSN
  • 0028-0836 (ISSN-L)
  • 1476-4687
OpenAlex
W4403680773
Cited by
132
References
24
Citations by year
2026: 63 · 2025: 62 · 2024: 7
FWCI
18.43
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
Funders
DeepMind
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
text watermarking, large language models, synthetic text detection, watermark detection, Gemini, responsible AI use

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Scalable watermarking for identifying large language model outputs. https://app.cochat.ai/papers/10.1038/s41586-024-08025-4
The paper
Dathathri, S., See, A., Ghaisas, S., Huang, P.-S., McAdam, R., Welbl, J., Bachani, V., Kaskasoli, A., Stanforth, R., Matejovicova, T., Hayes, J., Vyas, N., Merey, M. A., Brown-Cohen, J., Bunel, R., Balle, B., Cemgil, T., Ahmed, Z., Stacpoole, K., . . . Kohli, P. (2024). Scalable watermarking for identifying large language model outputs. Nature. https://doi.org/10.1038/s41586-024-08025-4

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.