Sakana AI is working on "defense and intelligence" as a focus area alongside finance. The aim is for AI to generate hypotheses from volumes of data that humans cannot read through, and for analysts to be able to use those hypotheses together with their supporting evidence.In March of this year, we collaborated with Yomiuri Shimbun to analyze narratives of China's criticism of Japan from 1.1 million posts on social media.In this analysis, the hypotheses generated by AI were corroborated through reporting by journalists.
At first glance, social media seems like a mirror reflecting public opinion, but simply counting the number of posts often misses a great deal. For example, a candidate who appeared to have strong support before an election may in fact have been supported only by mobilization within the party, or most of a sudden surge in reactions may in fact have been criticism. To grasp the true situation, it is necessary to read hundreds of thousands of statements while tracing back to "who wrote them and why."
Using AI makes it possible to conduct social media analysis that delves into the content of such posts. This time, as a case study, we analyzed the U.S. midterm elections, with voting to be held on November 3.
▶ Full report (PDF):AI Analysis of the U.S. Midterm Elections — Texas Senate Race
Narrative Analysis from Social Media
Sakana AI's narrative analysis system extracts discourse in units of narrative—"who did what, and how it is evaluated"—and measures its spread by the number of people who retold that narrative in their own words. The extracted narratives were checked against the original text by multiple AI models, and only those that were verified were kept in the report. For the approximately 710,000 reactions in this case as well, it estimates the ratio of support to criticism with a statistical margin of error.
Deep analysis of hypotheses
Looking only inside social media can sometimes make it impossible to notice distortions in the discourse space. So we had our product "Sakana Marlin", which conducts deep research through long-duration reasoning, investigate public information such as opinion polls, campaign finance, prediction markets, and news reports, and cross-checked it against observations on X. Marlin has a mechanism in which multiple AIs explore hypotheses in parallel and evaluate one another. By making full use of Marlin, it became possible to trace, with sources, whether change occurred first inside or outside social media.
What we learned in the Texas Senate race
In Texas, where Republicans have kept winning for about 30 years, this Senate race is expected to be close. Using AI as described above, we analyzed 2,042 posts on X by both candidates and both parties' state organizations, along with 705,976 reactions to them, and found, for example, the following.
- Much of the "support" directed at candidate Paxton before the Republican runoff was mobilization for the intra-party nomination contest. Once the runoff ended, the share of support fell from about 50% to just under 20%.
- What was exchanged in the reply sections of both candidates was a back-and-forth in which the two camps went into each other's sections to attack one another. Reading the content and trajectory of these reactions shows that candidate Talarico's relative advantage was indicated in June, and that as the main issues—candidate quality, the economy, religion, and data centers—shifted, the race became a narrowly contested one in the final stage.
- The discourse on X revealed complex movements not visible in opinion polls, such as shifts in evaluations of candidate Paxton and the narratives and assessments surrounding religion and data centers for both candidates. In addition, the shift in evaluations of both candidates appeared about two months before major rating organizations such as the Cook Political Report changed their judgments.
Toward intelligence conducted through collaboration with AI
It can be said that insights such as those above cannot be derived by humans or AI alone. Sakana AI recently announcedThe latest version of "The AI Scientist"Therefore, we proposed "Co-Discovery," in which humans and AI make scientific discoveries together, as the future of science. We believe the importance of this kind of human-AI collaboration is the same in the intelligence domain. AI derives hypotheses from information at a scale humans cannot read through and presents them together with supporting evidence. Humans who receive them examine the hypotheses, conduct interviews and additional observations, and then make judgments. Analyze using AI, derive insights, and leave the final judgment to humans. We believe this kind of collaboration between AI and humans is what intelligence should look like in the AI era.
Details of the analysis of the Texas Senate race are compiled in the full report.
▶ Full report (PDF):AI Analysis of the U.S. Midterm Elections — Texas Senate Race
Sakana AI
Sakana AI is looking for people who will open up the future of AI together with us. Our company'sApplication GuidelinesPlease take a look.
