DGIST Develops 'BREW' Technology to Identify AI-Generated Text

A DGIST research team has developed 'BREW', a technology capable of detecting text generated by artificial intelligence (AI). This technology proves whether AI generated the text and verifies its origin by embedding a unique, invisible digital watermark into AI-created text, enabling accurate identification of AI-written content and confirmation of its authenticity and source.

With the proliferation of generative AI, verifying the truthfulness and origin of content has become crucial. Existing technologies, however, were limited by their tendency to mistakenly flag human-written text as AI-generated. 'BREW' introduces a novel approach by dividing text into multiple segments for independent verification and tracking subtle sentence variations. It applies a 'window shifting' technique to ensure that the AI's authorship and origin can be identified even if the text is partially damaged or altered.

'BREW' maintained a detection rate of 96.5% even in a compromised environment where 10% of AI-generated text was altered with synonyms. The false positive rate, which misidentifies human-written text as AI-generated, was reduced to approximately 2%. The research on 'BREW' by Professor Kim Young-sik's team at DGIST has been accepted by the 'ICML 2026' conference and is scheduled to be presented at COEX in Seoul this July.