Tracing model-generated DNA with position-independent watermarking
Abstract
Genomic language models can write synthetic DNA that carries no record of its origin. A generation-time watermark could provide a provenance signal, if a verifier can detect it using only the DNA sequence, a key, and published detector settings, without knowing where the generated region starts, which strand it lies on, or how it was divided into six-base tokens. We embed the SynthID tournament watermark in two genomic language models, Carbon and GENERator-v2. The verifier searches both strands, every start position, and four window lengths, and corrects its decision for the whole search. In a detection cohort of 1,544 held-out prompts, it found all 3,088 marked sequences per model, unedited and after one substituted, inserted, or deleted base. For ordinary sequences, the one-sided 95% upper confidence bound on the false-positive rate was at most 0.850%; one bound for the wrong-key control reached 1.015% (Carbon, after a deletion). In a development cohort, neither model showed a detectable change in likelihood or in predefined sequence measures. Under random edits placed without reference to the detector, detection stayed complete or nearly complete up to a 2% per-base edit rate. These results show statistical detection, not biological function, secret-key security, or robustness to an editor who sees the detector.
Related
- Effective macrophage clearance of Klebsiella pneumoniae requires the inducible nitric oxide synthase iNOS and is independent of reactive oxygen species generated by NADPH oxidase
- Tracing the Evolutionary Histories of Human Cholinergic Proteins Across Eukaryotes: A Comparative Phylogenomic Study
- Tracing the origin of non-brittle rachis TtBTR1 alleles in wheat
- A Cre-responsive reporter quail for flexible genetic lineage tracing
- Tracing the dynamic recruitment trajectory of Optineurin between the centrosome and other macromolecular assemblies, signifying its role in cell cycle progression
BibTeX
@misc{provatas12026tracing,
title = {{Tracing model-generated DNA with position-independent watermarking}},
author = {Kimon Antonios Provatas1 and Christos Galanopoulos and Ilias Georgakopoulos-Soares},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.04763},
eprint = {2610.04763},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.04763},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Tracing model-generated DNA with position-independent watermarking
AU - Kimon Antonios Provatas1
AU - Christos Galanopoulos
AU - Ilias Georgakopoulos-Soares
PY - 2026
DA - 2026/10/03
PB - arXiv
DO - 10.48550/arxiv.2610.04763
UR - https://arxiv.org/abs/2610.04763
AB - Genomic language models can write synthetic DNA that carries no record of its origin. A generation-time watermark could provide a provenance signal, if a verifier can detect it using only the DNA sequence, a key, and published detector settings, without knowing where the generated region starts, which strand it lies on, or how it was divided into six-base tokens. We embed the SynthID tournament watermark in two genomic language models, Carbon and GENERator-v2. The verifier searches both strands, every start position, and four window lengths, and corrects its decision for the whole search. In a detection cohort of 1,544 held-out prompts, it found all 3,088 marked sequences per model, unedited and after one substituted, inserted, or deleted base. For ordinary sequences, the one-sided 95% upper confidence bound on the false-positive rate was at most 0.850%; one bound for the wrong-key control reached 1.015% (Carbon, after a deletion). In a development cohort, neither model showed a detectable change in likelihood or in predefined sequence measures. Under random edits placed without reference to the detector, detection stayed complete or nearly complete up to a 2% per-base edit rate. These results show statistical detection, not biological function, secret-key security, or robustness to an editor who sees the detector.
ER -