KAIST researchers analyzed about 110 million news comments accumulated over 20 years with artificial intelligence (AI) and developed technology that detects accounts and message patterns suspected of foreign-linked influence operations.
KAIST said on the 12th that a research team led by Graduate School of Culture Technology professor Lee Won-jae and School of Computing professors Cha Mi-young and Oh Hyeon jointly developed "explainable AI" technology with a team led by Torsten Holz at the Max Planck Institute in Germany that analyzes influence operations in online news comments and presents the grounds for its judgments.
Online influence operations refer to acts in which a specific group systematically spreads comments or posts to influence users' perceptions or public opinion. Existing AI technology can classify suspicious accounts, but it has had the limitation of being hard to explain which expressions or behaviors it based its judgment on.
The researchers started with 70 accounts that the Institute for National Security Strategy (INSS) had previously identified as foreign-linked accounts. They tracked accounts that were associated with these or repeatedly posted comments on the same articles and analyzed about 110 million comments posted on Naver News from 2006 to last year.
The AI examined whether the comments contained expressions or contexts that suggested foreign links, then analyzed whether emotions that could fuel conflict—such as moral condemnation or excessive praise—appeared. It then identified which countries, individuals, or groups those emotions targeted. It was also designed to present behavior patterns, including the comment expressions that formed the basis of its judgment along with the account's activity frequency and duration and connections with other suspicious accounts.
Among about 4 million Naver News users, the researchers classified 23,998 accounts as suspected of foreign-linked influence operations. The analysis found that, rather than consistently supporting a particular political camp, messages from these accounts were relatively more prominent in criticizing domestic politicians or Korean society and stoking conflict.
Seven of the top 10 targets with the most reactions were well-known domestic politicians. The targets were not concentrated on a particular party or ideology, spanning progressive and conservative camps, former and current presidents and presidential candidates, and political parties.
The researchers said foreign-linked influence operations should be assessed not only by whether they support a particular political force, but also by whether they expand existing social conflicts or foster distrust.
Lee Won-jae said, "The analysis found that suspicious accounts showed a pattern of criticizing Korea and domestic politicians and fueling conflict rather than directly praising foreign countries," adding, "It can be used to screen messages for priority review during times when social conflict grows, such as elections."
The study is scheduled to be presented on the 13th at the international computer security conference USENIX Security Symposium 2026.
References
arXiv (2026), DOI: https://doi.org/10.48550/arXiv.2606.22785