A research team led by Professor Kim Young-sik of the School of Electrical Engineering and Computer Science and the Artificial Intelligence major at DGIST said on the 2nd that it developed a text watermarking technology, "BREW (Block-wise Reliable Embedding for Watermarking)," that can verify the source of text generated by artificial intelligence (AI).
As the use of Generative AI expands, the need is growing for technologies that can identify who authored various texts such as news, documents, assignments, and creative works, but existing text watermarking methods have been criticized for the limitation of false positives that wrongly judge human-written text as AI-generated text.
BREW embeds an invisible digital watermark into AI-written text and is designed to verify whether content is AI-generated and its source even if sentences are partially edited or damaged afterward.
The research team applied a method that divides text into multiple blocks and verifies each independently. It also introduced a "window-shifting" technique to counter attempts to erase the watermark by replacing words or changing sentence structures. This technique restores disrupted alignment caused by sentence changes so the watermark can be traced again.
In experiments, BREW achieved a 96.5% detection rate even when 10% of the AI-generated text was replaced with synonyms. It maintained performance even for relatively short texts of around 200 words, and the false positive rate of wrongly judging human-written text as AI-generated text was as low as about 2%.
Kim said, "This study is meaningful in that it reduces the false positive problem of existing text watermarking technologies and increases traceability even when the text is partially modified," adding, "It could be used to verify the source of AI-generated content and protect digital copyrights."
The paper on this study was accepted by ICML 2026, a conference in the AI field, and the research team plans to present the results at ICML 2026 to be held at COEX in Seoul in Jul.
References
arXiv (2026), DOI: https://doi.org/10.48550/arXiv.2605.00348