Researchers at KAIST and Max Planck Institute have developed an artificial intelligence (AI) system that can detect signs of foreign-linked attempts to shape public opinion through news comment sections. The system examined 110 million comments posted on the Naver News platform between 2006 and 2025. It is designed to be explainable, meaning it can show the specific reasons and phrases behind each of its decisions rather than simply labeling accounts as suspicious without evidence.
An online influence operation is an organized effort in which certain accounts repeatedly post comments to steer public discussion in a chosen direction. Earlier detection tools often produced blacklists of accounts but could not clearly explain why those accounts were flagged. To address this gap, the researchers began with 70 accounts already identified by South Korea’s Institute for National Security Strategy as having foreign links. They then followed connected accounts and those that repeatedly commented on the same articles, building a large collection of data for analysis.
How the detection process works
The artificial intelligence examines each account in three stages. First it looks for clues that the writer may be connected to a foreign source. Second it checks whether the language contains strong polarizing expressions, such as harsh moral condemnation or uncritical praise. Third it identifies the country or person toward which that emotional language is directed. In addition, the system studies patterns of behavior, including how often an account posts, how long it has existed, and whether it interacts with other flagged accounts.
From roughly four million users, the system identified 23,998 accounts whose activity matched patterns of suspected opinion manipulation. These accounts did not consistently favor one political side. Instead they criticized figures and parties from both progressive and conservative camps. The apparent goal was to heighten internal conflict and distrust within Korean society rather than to promote any particular group. The most engaging targets were often well-known domestic political figures.
The researchers note that the technology could help platforms monitor comment sections during elections or national crises. They stress, however, that it should serve as a tool to assist human reviewers rather than automatically block accounts.
This research will be presented at the USENIX Security Symposium 2026.