Scalable watermarking for identifying large language model outputs
Open accessThe 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.
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
Cites
Statistics
Figures







Verification
Claims the paper makes
11 Valid · 4 Not valid · 12 Not checked · 60 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“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).”
- Also stated in: § SynthID-Text provides better detectability than existing watermarks · plain
“In the distortionary category, we compare the trade-off of detectability and text quality for distortionary SynthID-Text and Soft Red List. Both methods have a strength parameter that controls this trade-off; in Fig. 3c, we see that the trade-off is more favourable for distortionary SynthID-Text.”
§ Scalable watermarking for identifying large language model outputs > Evaluation > SynthID-Text provides better detectability than existing watermarks
What the sources say
- Supportedcites [23] Kirchenbauer, J. et al. A watermark for large language models. In Proc. 40th International Conference on Machine Learning 17061–17084 (PMLR, 2023). source
The superscript is attached to the noun phrase 'the distortionary Soft Red List watermark', so what is attributed to this source is the identification of that method and its distortionary nature. Our copy introduces exactly that: Section 3 / Algorithm 2 ('A more sophisticated watermark', pp. n/a) defines the soft red list rule that adds a constant delta to green-list logits, and Section 4.2 ('Impact on quality of generated text') plus the 'Strength vs Text Quality' paragraph (Figure 2, left) state that the soft watermark does impact perplexity for moderate-entropy tokens and plots the trade-off between watermark strength (z-score) and text quality (perplexity). The head-to-head comparison against SynthID-Text and the log(perplexity) axis are the citing paper's own measurement, not a claim this source makes — our copy never mentions SynthID (searched: 'SynthID', 'comparison with other watermarking methods'). (LLM, conf=0.82)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § SynthID-Text provides better detectability than existing watermarks · plain
Every source the paper cites for this claim was checked and supports it.
In the paper
“Furthermore, we provide an algorithm to combine generative watermarking with speculative sampling5—a frequently used technique to increase LLM text generation speed—allowing for the integration of SynthID-Text into large-scale production systems with negligible additional computational overhead.”
§ Scalable watermarking for identifying large language model outputs > Main
What the sources say
- Supportedcites [5] Chen, C. et al. Accelerating large language model decoding with speculative sampling. Preprint at https://arxiv.org/abs/2302.01318 (2023). source
The citation identifies speculative sampling as a technique for speeding LLM text generation. The Introduction presents it as an algorithm to accelerate transformer sampling, and the Conclusion reports speedups on benchmark tasks. The statement about integrating SynthID-Text is the citing authors’ own contribution, not a claim attributed to this source. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“To enable watermarking at scale, we develop an algorithm integrating watermarking with speculative sampling, an efficiency technique frequently used in production systems5.”
§ Scalable watermarking for identifying large language model outputs > Abstract
What the sources say
- Partly supportedcites [5] Chen, C. et al. Accelerating large language model decoding with speculative sampling. Preprint at https://arxiv.org/abs/2302.01318 (2023). sourceExperimental — accuracy not yet measured
The Introduction describes speculative sampling as a practical method for reducing decoding latency and reports a 2–2.5× speedup, supporting its description as an efficiency technique. The cited paper does not establish that it is frequently used in production systems. The sentence says the citing authors develop the watermarking algorithm; it does not attribute that development to this source. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Furthermore, we provide an algorithm to combine generative watermarking with speculative sampling5—a frequently used technique to increase LLM text generation speed—allowing for the integration of SynthID-Text into large-scale production systems with negligible additional computational overhead.”
§ Scalable watermarking for identifying large language model outputs > Main
What the sources say
- Supportedcites [5] Chen, C. et al. Accelerating large language model decoding with speculative sampling. Preprint at https://arxiv.org/abs/2302.01318 (2023). source
The Abstract identifies speculative sampling as the paper’s algorithm for accelerating transformer decoding and reports a 2–2.5× speedup. In the citing sentence, “we provide an algorithm to combine” watermarking with speculative sampling describes the citing authors’ contribution; citation 5 identifies the speculative-sampling technique, not the combination. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We find no significant difference between non-distortionary SynthID-Text and the equivalent unwatermarked model in terms of perplexity or performance on automated benchmarks.”
§ Scalable watermarking for identifying large language model outputs > Evaluation > SynthID-Text preserves quality including in a large-scale-production system
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
45 more claims
Create a free account to see all 87 claims and to verify the 3 references still open.
- It may state two separate claims as one. 27
- Probably not a substantial claim. 16
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?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Technology like watermarking could help improve the accountability of world model usage [30].”
From Understanding World or Predicting Future? A Comprehensive Survey of World Models (2026) · cites this paper as [30]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“A creator-centric how of AI attribution remains an open question – while algorithms for identifying and watermarking AI-generated content offer useful technological approaches [20, 28, 36], they may not encompass the extent and nuance of creators’ attribution needs.”
From Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co-Creation (2025) · cites this paper as [20]
Citations
Cites
A learning algorithm for Boltzmann machines
Ackley, D. H., Hinton, G. E., & Sejnowski, T. J. (1985) · Cognitive Science · doi:10.1207/s15516709cog0901_7 · Cited by 3,402
Hierarchical Neural Story Generation
Fan, A., Lewis, M., & Dauphin, Y. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1082 · Cited by 1,266
Hierarchical Neural Story Generation
Fan, A., Lewis, M., & Dauphin, Y. (2018) · arXiv · doi:10.48550/arxiv.1805.04833 · Cited by 1,242
Gemini: A Family of Highly Capable Multimodal Models
Gemini Team, Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. (2023) · arXiv · doi:10.48550/arxiv.2312.11805 · Cited by 990
Self-citationAI models collapse when trained on recursively generated data
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024) · Nature · doi:10.1038/s41586-024-07566-y · Cited by 793 · cited 1×
“Given the widespread use of LLMs in education, software development and web content generation, identification and attribution of LLM text is critical to ensure safe and responsible use of the technology 8–11 .”
MethodSelf-citationGPT detectors are biased against non-native English writers
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023) · Patterns · doi:10.1016/j.patter.2023.100779 · Cited by 673 · cited 1×
“In particular, they are known to perform poorly on out-of-domain data and may have higher false-positive rates for certain groups, such as non-native speakers 17 .”
BackgroundKöbis, N., & Mossink, L. D. (2021) · Computers in Human Behavior · doi:10.1016/j.chb.2020.106553 · Cited by 406
Elkhatat, A. M., Elsaid, K., & Almeer, S. (2023) · International Journal for Educational Integrity · doi:10.1007/s40979-023-00140-5 · Cited by 392 · cited 1×
“However, post hoc detection systems can themselves be computationally expensive to run, and their practical usage is limited by their inconsistent performance 16 .”
BackgroundHuman heuristics for AI-generated language are flawed
Jakesch, M., Hancock, J. T., & Naaman, M. (2023) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2208839120 · Cited by 342
Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. D. L., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al. (2023) · arXiv · doi:10.48550/arxiv.2310.06825 · Cited by 340 · cited 1×
“2 Mistral 7B-IT model 29 .”
Research works printed but not linked: 7
- 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).
- 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).
- 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).
- Christ, M., Gunn, S. & Zamir, O. Undetectable watermarks for language models. In Proc. Thirty Seventh Conference on Learning Theory 1125–1139 (PMLR, 2024).
- Jovanović, N., Staab, R. & Vechev, M. Watermark stealing in large language models. In Proc. 41st International Conference on Machine Learning 22570–22593 (PMLR, 2024).
- 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).
- 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
Generative AI in Marketing and Principles for Ethical Design and Deployment
Hermann, E., & Puntoni, S. (2025) · Journal of Public Policy & Marketing · doi:10.1177/07439156241309874 · Cited by 134
“…digital watermarking (i.e., embedding nonvi-sible but identi fi able markers to trace the origin of generated content and differentiate it from human-created content), risk management and veracity checks, and content restoration measures and processes (Dathathri et al. 2024; Epstein et al. 2023).”
Suchikova, Y., Tsybuliak, N., Teixeira da Silva, J. A., & Nazarovets, S. (2026) · Accountability in Research · doi:10.1080/08989621.2025.2544331 · Cited by 70
Understanding World or Predicting Future? A Comprehensive Survey of World Models
Ding, J., Zhang, Y., Shang, Y., Zhang, Y., Zong, Z., Feng, J., Yuan, Y., Su, H., Li, N., Sukiennik, N., et al. (2026) · ACM Computing Surveys · doi:10.1145/3746449 · Cited by 169
“Technology like watermarking could help improve the accountability of world model usage [41].”
Khosravi, B., Purkayastha, S., Erickson, B. J., Trivedi, H., & Gichoya, J. W. (2025) · The Lancet Digital Health · doi:10.1016/j.landig.2025.100890 · Cited by 62
Ullah, M., Naeem, S. B., & Boulos, M. N. K. (2024) · Big Data and Cognitive Computing · doi:10.3390/bdcc8120194 · Cited by 77
Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co-Creation
He, J., Houde, S., & Weisz, J. D. (2025) · Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3713522 · Cited by 36
The sociolinguistic foundations of language modeling
Grieve, J., Bartl, S., Fuoli, M., Grafmiller, J., Huang, W., Jawerbaum, A., Murakami, A., Perlman, M., Rœmling, D., & Winter, B. (2025) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2024.1472411 · Cited by 87
Large Language Models in Computer Science Classrooms: Ethical Challenges and Strategic Solutions
Azoulay, R., Hirst, T., & Reches, S. (2025) · Applied Sciences · doi:10.3390/app15041793 · Cited by 47
Mohammadabadi, S. M. S., Kara, B. C., Eyüpoğlu, C., Uzay, C., Tosun, M. S., & Karakuş, O. (2025) · Electronics · doi:10.3390/electronics14183580 · Cited by 90
SoK: Watermarking for AI-Generated Content
Zhao, X., Gunn, S., Christ, M., Fairoze, J., Fábrega, A., Carlini, N., Garg, S., Hong, S., Nasr, M., Tramèr, F., et al. (2025) · IEEE Symposium on Security and Privacy (S&P) · doi:10.1109/sp61157.2025.00178 · Cited by 134
“SynthID [57, 58] is a pio-neering technology developed by Google to address the challenge of identifying AI-generated content across various media formats, including text, audio, images, and video.”
Deepfakes on Demand: The rise of accessible non-consensual deepfake image generators
Hawkins, W., Mittelstadt, B., & Russell, C. (2025) · Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3715275.3732107 · Cited by 20
“This has led to concerns relating to misinformation, efforts to prevent generation of identifiable individuals, and watermarking of AI-generated content to track synthetic media [10, 25].”
Liu, C., Tang, D., Zhu, H., Wang, L., Cai, Q., & Nie, Q. (2025) · Journal of Manufacturing Systems · doi:10.1016/j.jmsy.2025.12.016 · Cited by 28
Watermarking for Large Language Models: A Survey
Yang, Z., Zhao, G., & Wu, H. (2025) · Mathematics · doi:10.3390/math13091420 · Cited by 45
Ng, J. Y., Krishnamurthy, M., Deol, G., Al-Khafaji, W. A.-Z., Balaji, V., Abebe, M., Adhvaryu, J., Karrthik, T., Mohanakanthan, P., Vellaparambil, A., et al. (2025) · Research Integrity and Peer Review · doi:10.1186/s41073-025-00178-8 · Cited by 14
“Addressing these challenges proactively will ensure that AICs serve as a valuable complement to, rather than a replacement for, human editorial expertise.”
Advancements in AI-Generated Content Forensics: A Systematic Literature Review
Xu, Q., Mu, W., Li, J., Sun, T., & Jiang, X. (2025) · ACM Computing Surveys · doi:10.1145/3760526 · Cited by 26
“Google’s system [89] represents a major step towards practical deployment, offering a production-ready solution with minimal latency and high accuracy.”
Trustworthy AI for Whom? GenAI Detection Techniques of Trust Through Decentralized Web3 Ecosystems
Calzada, I., Németh, G., & Al-Radhi, M. S. (2025) · Big Data and Cognitive Computing · doi:10.3390/bdcc9030062 · Cited by 16
AI Rules? Characterizing Reddit Community Policies Towards AI-Generated Content
Lloyd, T., Gosciak, J., Nguyen, T., & Naaman, M. (2025) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3713292 · Cited by 42
The impact of advanced AI systems on democracy
Summerfield, C., Argyle, L. P., Bakker, M. A., Collins, T., Durmus, E., Eloundou, T., Gabriel, I., Ganguli, D., Hackenburg, K., Hadfield, G. K., et al. (2025) · Nature Human Behaviour · doi:10.1038/s41562-025-02309-z · Cited by 87
“However, many current techniques are quite easy to circumvent 75 , and researchers acknowledge that watermarking is not a silver bullet for avoiding misinformation 76 .”
Adaptive Robust Watermarking for Large Language Models via Dynamic Token Embedding Perturbation
Zeng, Z., Lin, H., Zhang, S., & Wang, B. (2026) · IEEE Access · doi:10.1109/access.2026.3653833 · Cited by 33
“More recently, Google’s SynthID-Text [32] demonstrated the feasibility of watermarking at production scale with speculative sampling integration, achieving high detection accuracy with minimal latency overhead.”
Attack and defense techniques in large language models: A survey and new perspectives
Liao, Z., Chen, K., Lin, Y., Li, K., Liu, Y., Chen, H., Huang, X., & Yu, Y. (2025) · Neural Networks · doi:10.1016/j.neunet.2025.108388 · Cited by 51
Improving authenticity and provenance in digital biomarkers: the case for digital watermarking
Mahajan, A., & Powell, D. (2025) · npj Digital Medicine · doi:10.1038/s41746-024-01374-4 · Cited by 9
The Purity Myth: Why Stigmatizing GAI in Academic Writing Is Harmful
Suchikova, Y., & Tsybuliak, N. (2025) · Science Communication · doi:10.1177/10755470241313233 · Cited by 10
“Watermarks aim to ensure accountability without compromising quality (Dathathri et al., 2024; Gibney, 2024).”
Bisztray, T., Cherif, B., Dubniczky, R. A., Gruschka, N., Borsos, B., Ferrag, M., Kovács, A., Mavroeidis, V., & Tihanyi, N. (2025) · AISec@CCS · doi:10.1145/3733799.3762964 · Cited by 24
AI detecting AI in academic writing: Why most AI detector findings are false
Tsigaris, P., & Teixeira da Silva, J. A. (2026) · Next Research · doi:10.1016/j.nexres.2026.101396 · Cited by 11
Zhang, Z., Jin, R., Liu, J., Xu, G., Wang, X., Žitnik, M., Zhao, H., Cong, L., & Wang, M. (2024) · bioRxiv (Cold Spring Harbor Laboratory) · doi:10.1101/2024.10.23.619960 · Cited by 9
“Although watermarks have been applied for Language models [ 33, 34, 35, 36, 37, 38 ] (e.g., SynthID-Text [ 37 ]) and text-to-image models [ 39,40,41,42 ] (e.g., AquaLoRA [ 41 ]), extending these methods to protein structure data presents unique challenges.”
Navigating Generative AI Disclosure, Ownership, and Accountability in Co-Creative Domains
Do, H. J., Feldman, M. Q., He, J., Hwang, A. H.-C., & Kim, S. (2025) · Symposium on Human-Computer Interaction for Work · doi:10.1145/3707640.3729212 · Cited by 8
The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems
Staufer, L., Feng, K. J. K., Wei, K., Bailey, L., Duan, Y., Yang, M., Ozisik, A. P., Casper, S., & Kolt, N. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3806728 · Cited by 71
“Agent builder platforms (Zapier, Salesforce, OpenAI AgentKit) generally provide optional guardrail modules and little information on built-in protections.”
Artificial intelligence content detection techniques using watermarking: A survey
Kumar, N., & Singh, A. K. (2025) · Image and Vision Computing · doi:10.1016/j.imavis.2025.105728 · Cited by 8
Using Generative AI in Historical Practice
Fox, Y. (2026) · Cambridge University Press eBooks · doi:10.1017/9781009740067 · Cited by 6
Debiasing Watermarks for Large Language Models via Maximal Coupling
Xie, Y., Li, X., Mallick, T., Su, W., & Zhang, R. (2025) · Journal of the American Statistical Association · doi:10.1080/01621459.2025.2520455 · Cited by 34
Jones, C., & Bergen, B. K. (2026) · Artificial Intelligence Review · doi:10.1007/s10462-026-11517-6 · Cited by 61
Yi, X., Li, Y., Zheng, S., Wang, L., Wang, X., & He, L. (2025) · Knowledge-Based Systems · doi:10.1016/j.knosys.2025.114295 · Cited by 16
Google unveils invisible ‘watermark’ for AI-generated text
Gibney, E. (2024) · Nature · doi:10.1038/d41586-024-03462-7 · Cited by 3
AI watermarking must be watertight to be effective
(2024) · Nature · doi:10.1038/d41586-024-03418-x · Cited by 3
Mihăilescu, E., & Chiper, D. F. (2025) · International Symposium on Signals, Circuits and Systems · doi:10.1109/isscs66034.2025.11105620 · Cited by 3
“For instance, recent work in 2024 introduced a scalable watermarking method for Large Language Model outputs, embedding statistically detectable patterns in AI-generated text without altering its readability [1].”
InfluentialFang, X., Hu, P. J., Chau, M., & Chen, H. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5455094 · Cited by 3
Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text
Cheng, Y., Sadasivan, V. S., Saberi, M., Saha, S., & Feizi, S. (2025) · Advances in Neural Information Processing Systems 38 · doi:10.52202/085713-1588 · Cited by 3
Tracing the Pen: Electronic Health Records Amid the Rise of Generative AI
Nargesi, A. A., You, J. G.-T., Bitterman, D. S., Succi, M. D., Mishuris, R. G., Poon, E. G., & Landman, A. (2026) · npj Digital Medicine · doi:10.1038/s41746-026-02508-6 · Cited by 4
“Suchmethods,albeitsub-jecttosecurityvulnerabilities,arenowpubliclyavailableandrepresenta scalableapproachwithminimalcomputationalcost 71 .”
Trust or distrust? AIGC trustworthiness and an extended analysis within nomology framework
Li, M., Du, R., Ai, S., Hunt, R. A., & Xie, J. (2025) · Data Science and Management · doi:10.1016/j.dsm.2025.09.005 · Cited by 4
Ng, J. Y., Krishnamurthy, M., Deol, G., Al-Khafaji, W. A.-Z., Balaji, V., Abebe, M., Adhvaryu, J., Karrthik, T., Mohanakanthan, P., Vellaparambil, A., et al. (2025) · doi:10.1101/2025.05.26.25328101 · Cited by 14
Suchikova, Y. (2025) · NanoEthics · doi:10.1007/s11569-025-00477-w · Cited by 5
“In the academic domain, concerns about transparency and accountability are also reflected in current debates over watermarking and disclosure of AI-generated content [48–51].”
In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?
Bucknall, B., Siddiqui, S., Thurnherr, L., McGurk, C., Harack, B., Reuel, A., Paskov, P., Mahoney, C., Mindermann, S., Singer, S., et al. (2025) · SuperIntelligence - Robotics - Safety & Alignment · doi:10.70777/si.v2i5.15509 · Cited by 5
Robust detection of watermarks for large language models under human edits
Li, X., Ruan, F., Wang, H., Long, Q., & Su, W. (2025) · Journal of the Royal Statistical Society Series B (Statistical Methodology) · doi:10.1093/jrsssb/qkaf056 · Cited by 29
Auditing unauthorized training data from AI generated content using information isotopes
Qi, T., Yin, J., Cai, D., Xie, Y., Wang, H., Hu, Z., Yang, P., Nan, G., Zhou, Z., Wu, C., et al. (2026) · Nature Communications · doi:10.1038/s41467-026-68862-x · Cited by 4
Llorca, R. E. (2024) · Cuadernos de Derecho Privado · doi:10.62158/cdp.69 · Cited by 9
SEFD: Semantic-Enhanced Framework for Detecting LLM-Generated Text
He, W., Hou, B., Shang, T., Tarzanagh, D. A., Long, Q., & Shen, L. (2024) · IEEE International Conference on Big Data (BigData) · doi:10.1109/bigdata62323.2024.10825110 · Cited by 11
Governing Framework for the Responsible Democratization of Large-Language Models in Banking
Mandal, A., Ramanayake, R., O'Neill, O., Flanagan, W., Kemal, N., Yekrangi, M., Chatbri, H., & Martin, C. (2025) · IEEE Internet Computing · doi:10.1109/mic.2025.3621976 · Cited by 5
Casal, J. E., Stewart, C. M., & Windsor, A. J. (2025) · PLoS ONE · doi:10.1371/journal.pone.0324611 · Cited by 2
Robustness Assessment and Enhancement of Text Watermarking for Google’s SynthID
Han, X., Li, Q., Ni, J.-B., & Zulkernine, M. (2025) · International Conference on Trust, Security and Privacy in Computing and Communications · doi:10.1109/trustcom66490.2025.00109 · Cited by 9
“Among existing approaches, Google DeepMind’s SynthID-Text is state-of-the-art [3], notable as the only watermarking method integrated into a real-world product (Google’s Gemini models), a rare industrial deployment in this domain.”
InfluentialWatermarking LLMs — Challenges and Opportunities in Electronic Design Automation
DeLorenzo, M. C., Tieu, P., Chen, C., Gohil, V., & Rajendran, J. (2025) · IEEE International Conference on Omni-layer Intelligent Systems (COINS) · doi:10.1109/coins65080.2025.11125763 · Cited by 3
Adiwibowo, A. S., & Wahyuni, E. T. (2026) · Information Technology and Management · doi:10.1007/s10799-026-00488-3 · Cited by 1
Proactive Schemes: A Survey of Adversarial Attacks for Social Good
Asnani, V., Yin, X. P., & Liu, X. (2026) · International Journal of Computer Vision · doi:10.1007/s11263-025-02631-z · Cited by 6
Character-Level Adversarial Samples Generation Method Based on Dual-Layer Text Watermark
Wu, H., Li, X., Zhang, S., & Zhu, H. (2025) · IEEE Access · doi:10.1109/access.2025.3594727 · Cited by 2
“Notably, while modern generative language models have incorporated text watermarking for content provenance [8], [9], they remain susceptible to adversarial attacks, revealing a fundamental connection between these two technical paradigms in the adversarial-defense landscape.”
PGmark : Enhancing text quality in language model watermarking via probability guidance
Jiang, Z. Q., Wang, H., Hou, Q., & Jiao, R. (2026) · Expert Systems with Applications · doi:10.1016/j.eswa.2026.131120 · Cited by 3
Innamark: A Whitespace Replacement Information-Hiding Method
Hellmeier, M., Norkowski, H., Schrewe, E.-C., Qarawlus, H., & Howar, F. M. (2025) · IEEE Access · doi:10.1109/access.2025.3583591 · Cited by 3
“These ideas were integrated into SynthID, the watermarking engine used by Google’s Gemini LLM [16].”
LLM Unlearning on Noisy Forget Sets: A Study of Incomplete, Rewritten, and Watermarked Data
Wang, C., Zhang, Y., Wei, D., Jia, J., Chen, P.-Y., & Liu, S. (2025) · ACM Workshop on Artificial Intelligence and Security (AISec) · doi:10.1145/3733799.3762973 · Cited by 4
“For KGW, to-kens highlighted in red belong to the red list, and those in green belong to the green list; a higher proportion of red tokens reflects a stronger watermark signal.”
Watermarking Fine-Tuning Datasets for Robust Provenance
Gergov, I., & Tsochev, G. (2025) · Applied Sciences · doi:10.3390/app151910457 · Cited by 2
Yang, L., & Long, T. (2026) · International Journal of Digital Law and Governance · doi:10.1515/ijdlg-2026-0010 · Cited by 2
“Watermarking, metadata and machine-readable marking can make AI-generated outputs technically detectable and support subsequent moderation or veri fi cation (Dathathri et al. 2024; Liu et al. 2024).”
Function-preserving watermarking of AI-generated proteins
Stutz, D., Cowen-Rivers, A. I., Ortiz-Jiménez, G., Ratcliff, J., Zambaldi, V., Willmore, L., Abramson, J., Patani, H. H., Kouridi, C., Stimberg, F., et al. (2026) · Nature · doi:10.1038/s41586-026-10965-y · Cited by 4
Self-citationYet Another Watermark for Large Language Models
Bao, S., Shi, Y., Yang, Z., Wu, H., & Zhang, X. (2025) · International Conference on Communications and Broadband Networking (ICCBN) · doi:10.1109/iccbn67217.2025.11448000 · Cited by 2
“Another idea is to refine the sampling strategy for text generation, either at the token level [13] or sentence level [14], which does not require adjustments to model training, but would increase the inference complexity.”
Bridging Today and the Future of Humanity: AI Safety in 2024 and BeyondPreprint
Han, S. (2025) · Qeios · doi:10.32388/zhm9uq · Cited by 5
Zepf, S., & Colley, M. (2025) · INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION · doi:10.1145/3716553.3750736 · Cited by 1
“Risks: AI-enhanced presence and expression carry substantial risks, primarily concerning authenticity [18] and social connection.”
Guo, Z., Deng, Y., Liang, C., Zhang, H., Li, D., & Huang, M. T. (2025) · International Journal of Surgery · doi:10.1097/js9.0000000000004034 · Cited by 2
Black-Box Text Watermarking Algorithm towards Large Language Models
姚, 奕. (2025) · Modeling and Simulation · doi:10.12677/mos.2025.144276
LLM-Oriented Information Retrieval: A Denoising-First Perspective
Dai, L., Sun, L., Cao, F., Rao, Z., Yang, C., Liu, H., & Xiong, H. (2026) · International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) · doi:10.1145/3805712.3808544 · Cited by 2
From AI-generated content to agentic action: Security and safety threats in generative AI
Zhang, Z., Li, Q., Cao, J., Liu, L., & Ni, J. (2026) · Journal of Information and Intelligence · doi:10.1016/j.jiixd.2026.05.002 · Cited by 4
“…Low Moderate (long text required) Stable Signature [38] Decoder fine-tuning Multi-bit Low > 90% at 90% crop Tree-Ring [134] Initial noise perturbation 1 bit Zero JPEG, rotation, crop SynthID-Image [32] DNN encoder-decoder Multi-bit Low > 99% common edits Image StegaStamp [123] Encoder-decoder”
In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?
Bucknall, B., Siddiqui, S., Thurnherr, L., McGurk, C., Harack, B., Reuel, A., Paskov, P., Mahoney, C., Mindermann, S., Singer, S., et al. (2025) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3715275.3732201 · Cited by 3
Optimized Couplings for Watermarking Large Language Models
Long, C. X., Tsur, D., Verdun, C. M., Hsu, H., Permuter, H. H., & Calmon, F. P. (2025) · IEEE International Symposium on Information Theory (ISIT) · doi:10.1109/isit63088.2025.11195457 · Cited by 11
“S can be, for instance, generated from the hash of previous tokens in a sequence [5] or sophisticated tournament-like sampling strategies [12].”
Trustworthy AI for Whom in Smart Cities?
Calzada, I. (2025) · Studies in digital politics and governance · doi:10.1007/978-3-032-11888-2_3
Semantic-aware watermarking with adaptive injection for large language models
Zhao, J., Yang, H., Dong, H., He, H., Zhang, W., & Xu, H. (2026) · Knowledge-Based Systems · doi:10.1016/j.knosys.2026.116904
Dualmark: A novel dual watermarking approach for large language models
Qiang, Z., Hao, J., Qiang, J., Zhu, Y. Q., Zhang, C., Liu, Y., & Li, W. (2026) · Information Processing & Management · doi:10.1016/j.ipm.2026.104707 · Cited by 2
CODE ACROSTIC: Robust Watermarking for Code Generation
Li, L., Xin, S., Cao, Y., & Cao, X. (2026) · Lecture notes in computer science · doi:10.1007/978-981-95-7251-9_35
Gao, Y., Huang, T., Jiang, Y., Qu, K., Zhou, Z., & Shi, F. (2026) · IET Computers & Digital Techniques · doi:10.1049/cdt2/4161832
PromptHash: Robust Instruction Watermarks Against Paraphrase and Splicing in LLM Forensics
Ge, Z. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11463913 · Cited by 2
RINGet: A Robust Watermarking Framework for Diffusion-Based Video Generation
Yang, J., Zhang, W., Yang, L., Zeng, S., & Gao, T. (2025) · IEEE Transactions on Circuits and Systems for Video Technology · doi:10.1109/tcsvt.2025.3643532 · Cited by 2
“For instance, Dathathri et al. [48] proposed SynthID Text, a text watermarking scheme that preserves LLM functionality while enabling reliable detection through speculative sampling, achieving low latency and high accuracy.”
Angelier, V. A. (2026) · AI and Ethics · doi:10.1007/s43681-026-01376-w · Cited by 2
Open-Set Zero-Day Release Authorship Attribution of Code for Digital Forensics
Gonçalves, A., Carlson, I., Silva, P. C. D. L. E., Barros, P. H., & Ramos, H. S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-39895-6_18
Semantic-Oriented Robust Sentence Level Text Watermark for Large Language Models
Li, B., Zhang, K., Song, C., Chen, X., Zhou, H., & Hong, R. (2026) · IEEE Transactions on Computational Social Systems · doi:10.1109/tcss.2026.3687267
“SynthID [22] divides the vocabulary into multiple groups and uses tournament sampling to enable groups to compete for choosing the optimal next token, resulting in improving the quality of the generated watermarked text.”
Data Integrity for Machine Unlearning
Wang, C., & Zhang, Y. (2026) · Synthesis lectures on computer vision · doi:10.1007/978-3-032-17282-2_6
Shao, J., Liu, X., Zhang, Z., Liao, S., Wu, J., Yang, H., Zhao, Z., Wang, L., Liang, S.-F., Wang, X., et al. (2026) · Cell Reports Medicine · doi:10.1016/j.xcrm.2026.102926
Low-resource single-pass paraphrasing attacks on AI-generated text detection
Rydell, F., Hansen, P., Kanestad, L., Ljungar, A., & Öberg, S. (2026) · Intelligent Systems with Applications · doi:10.1016/j.iswa.2026.200710
Sicherheitsanforderungen an KI-Systeme
Halvani, O., & Müller, L. (2025) · Datenschutz und Datensicherheit - DuD · doi:10.1007/s11623-025-2092-5
“Eine von Google entwickelte Lösung mit dem Namen SynthID-Text [11] ermöglicht es, Texte, die von einem LLM generiert werden, mit einem Wasserzeichen zu versehen.”
Postmodern epistemik galat paradoksu: Üretken yapay zekâ kullanımı için bir etiketleme çerçevesi
Öncü, S. E., Bozkurt, A., Elibol, H., & Özer, F. (2026) · Açıköğretim Uygulamaları ve Araştırmaları Dergisi · doi:10.51948/auad.1922925
Bi-mark: A copyright marking framework for AI-generative-text
Yu, B., & Liu, G. (2025) · International Conference on High Performance Compilation, Computing and Communications (HP3C) · doi:10.1145/3774949.3774970
Modern Watermarking for Generative AI: Methods, Robustness, and Evaluation Metrics
Mehta, T., Thakkar, A., & Kadam, A. (2026) · Lecture Notes in Networks and Systems · doi:10.1007/978-3-032-27754-1_7
A Systematic Review and Comparative Analysis of AI-Generated Text Detection in Education
Daoudi, H., Riad, M., Salhi, I., Gouraguine, S., & Qbadou, M. (2026) · International Conference on Circuit, Systems and Communication (ICCSC) · doi:10.1109/iccsc71566.2026.11650035
TransMark: Lossless high-capacity watermarking for LLMs via neuron permutation invariances
Ye, P.-G., Chen, Z., Sang, H., & (207297), J. Z. (2026) · Computer Standards & Interfaces · doi:10.1016/j.csi.2026.104180
Deng, X., Zhou, T., Min, D., Zhu, G., Yu, K., & Yu, Y. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11462181
Ownership-Protected Semantic Communication via Signal Processing-Driven Robust Watermark
Han, B., Yang, X., Li, G., et al. (2026) · ICC 2026 - IEEE International Conference on Communications · doi:10.1109/icc59461.2026.11587337
A Training-free Method for LLM Text AttributionPreprint
Radvand, T., Abdolmaleki, M., Mostagir, M., & Tewari, A. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6169527 · Cited by 7
MAF-Detect: A Multi-Scale Adaptive Fusion Framework for Zero-Shot Detection of LLM-Generated Text
Bai, Z., & Zheng, Y. L. (2025) · International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI) · doi:10.1109/mlbdbi67855.2025.11331454 · Cited by 1
“Recent research [7] has begun to focus on the robustness of watermarking technology against various rewriting attacks and proposed evaluation and detection methods such as "WinMax"[8-9].”
Generating the Past: How Artificial Intelligence Summaries of Historical Events Affect Knowledge
Karell, D., Shu, M., Davidson, T., & Okura, K. (2025) · Social Science Computer Review · doi:10.1177/08944393251409744 · Cited by 2
Ng, J. Y., Bhavsar, D., Krishnamurthy, M., Dhanvanthry, N., Fry, D., Kim, J. W., King, A., Lai, J., Makwanda, A., Olugbemiro, P., et al. (2026) · Integrative Medicine Research · doi:10.1016/j.imr.2026.101403 · Cited by 1
Jiao, J., Zhou, J., Wu, C., et al. (2026) · Energy · doi:10.1016/j.energy.2026.142058
Toward Transparency of Synthetic Media: EU Regulatory Perspective and Its Implications
Łabuz, M. (2026) · Future of business and finance · doi:10.1007/978-3-032-07605-2_22
Why it’s so hard to tell if a piece of text was written by AI – even for AI
Tewari, A. (2025) · The Conversation · doi:10.64628/aai.ksptdekuu
Synthetic Media: Deepfakes, AI-Generated Content, and Authenticity in the Digital Society
Voinea, D. V. (2026) · Encyclopedia · doi:10.3390/encyclopedia6080175 · Cited by 1
Artificial Intelligence in Wealth Management
Herold, S., & Goltzsche, C. (2025) · Sustainable finance · doi:10.1007/978-3-032-02530-2_9
AGMark: Attention-Guided Dynamic Watermarking for Large Vision-Language Models
Li, Y., Yi, X., Shi, D., Cui, Y., de Melo, G., & Wang, L. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3817700 · Cited by 3
Rose, R. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6879243
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
- ORCID
- Venue
- Nature
- Published
- 2024-10-23
- Volume
- 634
- Issue
- 8035
- Pages
- 818-823
- ISSN
- 0028-0836 (ISSN-L)
- 1476-4687
- PMC
- PMC11499265
- 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
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- 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
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.