Sakana AI finds primary mobilization behind Paxton’s early support on X

A surge of supportive replies can tell the wrong election story. In a new Texas Senate analysis, Sakana AI says much of Ken Paxton’s early support on X reflected Republican primary mobilization. The company examined 2,042 posts and 705,976 responses, using AI to distinguish the arguments people made rather than simply count their reactions.

Reading the argument, not just the activity

The sample covered posts from the two candidates and their state party organizations, plus responses to those posts. Sakana’s system extracted narratives: accounts of who did something, what they did and how that action was judged. It measured a narrative’s spread by counting people who retold it in their own words.

That puts the content of a response at the center of the analysis. A busy reply section might contain backing for a candidate, criticism of that candidate or attacks from the opposing camp. Sakana says both candidates’ reply sections featured supporters crossing into the other side’s posts to attack.

The company describes a checking stage before those narratives entered the report. Multiple AI models compared the extracted accounts with the original posts, and only narratives that passed those checks were retained. For the roughly 706,000 responses, the system also estimated the proportions expressing support or criticism, with statistical margins of error.

Sakana then used its research product, Sakana Marlin, to look beyond X. Marlin examined polls, campaign finance, prediction markets and news, comparing that public information with the social-media observations. Its multiple AI systems explore hypotheses in parallel and evaluate one another’s work, according to the company.

A primary signal gives way to a close contest

The runoff marked a sharp change in the company’s reading of Paxton’s responses. Support fell from about half before the Republican runoff to less than one-fifth afterward. Those figures describe the analyzed reactions on X, not the share of Texas voters supporting him. Sakana’s explanation is that much of the earlier activity served the nomination fight.

The analysis identified a relative advantage for Talarico in June, followed by a narrow contest toward the end of the period studied. The arguments also changed: candidate qualities, the economy, religion and data centers became leading issues at different points. Sakana presents that changing mix as part of the election picture that simple response totals would miss.

An earlier signal still needs judgment

The study’s strongest timing claim is that changes in how the candidates were evaluated appeared on X about two months before major rating organizations, including Cook Political Report, revised their assessments. That is Sakana’s comparison of the sequence of changes—not a demonstrated election prediction. The finding concerns when signals emerged in different sources.

Sakana frames the work as intelligence produced through human-AI collaboration. AI generates hypotheses from more material than a person could read and presents supporting evidence. Human analysts then examine those hypotheses, conduct interviews or make additional observations, and reach a judgment. In that proposed workflow, the machine’s role is to surface leads with evidence, while responsibility for the conclusion remains with people.