Intelligence Report: US Midterm Elections seen from AI analysis of SNS

Sakana AI is working on "Defense/Intelligence" as a focus area alongside finance. The goal is for AI to create hypotheses from amounts of data that humans cannot read entirely, so that analysts can use those hypotheses along with evidence. In March of this year, in collaboration with Yomiuri Shimbun, they analyzed narratives of anti-Japan criticism by China from 1.1 million posts on SNS.

As a case study, they analyzed the US midterm elections where voting will be held on November 3. Sakana AI's narrative analysis system extracts discourse in units of stories such as "who did what and how it is evaluated," and multiple AI models cross-check these with the original text. Additionally, they used their product "Sakana Marlin" to perform deep research by cross-referencing observations on X with public information such as opinion polls and election funds.

In the analysis of the Texas Senate Election, the company analyzed 2,042 posts to X by both candidates and both parties' state organizations, along with 705,976 reactions. The analysis found that much of the "support" directed toward candidate Paxton before the Republican primary was mobilization toward the intra-party nomination battle, with the proportion of support dropping from about 50% to less than 20% after the primary ended. It also revealed an exchange where both camps went to each other's reply sections to attack one another.