The use of artificial intelligence (AI) technology as a tool for manipulating public opinion is on the rise. Although technologies for detecting AI-generated text are being developed, they primarily rely on lengthy, standardized texts in English. This means there are limitations in identifying AI comments on news platforms that are problematic in Korea.
The Korea Advanced Institute of Science and Technology (KAIST) and the National Security Research Institute announced on the 23rd that they have developed technology for detecting AI-generated Korean comments for the first time in the world. They improved the detection rate of AI-generated comments by analyzing the subtle differences between comments generated by AI and those written by humans.
Recently, generative AI technology is being mobilized for manipulating comments on news platforms. AI can even adjust emotions and tone to fit the context of news articles. A KAIST research team conducted a comparison to determine whether individuals could distinguish between AI-generated comments and those written by humans, finding that 67% of AI-generated comments were mistaken for being written by humans.
A larger problem is the low expense. According to OpenAI's generative AI, the cost of creating a single comment using the GPT-4o API is about 1 won. The major news platforms in Korea receive an average of 200,000 comments daily. This implies that for 200,000 won, the entire comment section of these news platforms could be manipulated. As long as one has their own graphics processing unit (GPU) infrastructure, it is practically possible to generate large numbers of comments for free.
It is not that there is no technology for detecting AI-generated text. However, the technologies that have been developed so far are based on lengthy texts written in English and are challenging to apply to short comments in Korean. Additionally, short comments have insufficient statistical features, and they often contain informal spoken expressions such as emojis or slang, making it difficult for existing detection models to function properly. The lack of data on comments generated by AI in Korean poses a significant issue.
The research team gathered data using 14 types of large language models (LLMs) trained on a vast amount of text information. Based on this, they constructed a dataset of Korean comments that mimics the style of actual users. The analysis revealed that AI-generated comments exhibit distinct linguistic patterns different from those of humans.
For instance, AI demonstrated a tendency to use formal expressions such as 'seems like' and 'regarding,' along with a high usage rate of connectives. In contrast, humans used more formatting characters like line breaks and multiple spaces. While 26% of comments written by humans included such formatting characters, only 1% of AI-generated comments did. The usage rate of repeated characters (e.g., ㅋㅋㅋㅋ, ㅎㅎㅎㅎ, etc.) was also significantly higher at 52% for human-written comments compared to 12% for AI-generated comments.
In terms of special character usage, AI predominantly employs standardized emojis recognized worldwide, whereas humans utilize a variety of characters rich in cultural specificity, such as Korean consonants (ㅋ, ㅠ, ㅜ, etc.) or special symbols (ㆍ, ♡, ★, •, etc.).
The 'XDAC' developed by KAIST precisely reflects these differences to enhance detection performance. It employs methods to transform formatting characters like line breaks and spaces and converts repeated character patterns into a form understandable by machines. It has also become possible to identify which AI model generated a comment by recognizing the unique linguistic characteristics of each LLM.
Thanks to these optimizations, XDAC was able to identify AI-generated comments with a probability of 98.5% based on the F1 score. The probability of identifying the comment-generating LLM was also 84.3%. The F1 score is a comprehensive performance metric that considers both accuracy and recall, used for evaluating AI models.
The research team expects that the XDAC detection technology will enable responses to AI-based manipulation of public opinion. Go Woo-young, a senior researcher at KAIST, noted, "This research is the world's first technology capable of accurately detecting short comments written by generative AI and identifying the generative model, which is significant in establishing a technological foundation for responding to AI-based manipulation of public opinion."
The results of this study were accepted at the main conference of the Association for Computational Linguistics (ACL) 2025, the leading academic conference in the field of artificial intelligence natural language processing. ACL 2025 will be held on July 27.