Polling averages have become the dominant tool for tracking elections, but they tell us who is ahead, not why. That's the argument from Jon Cohen, a former polling director at The Washington Post and Pew Research Center, who now runs an AI-assisted research firm. Writing in a new analysis, Cohen revisits his 2006 call for "standards, not averages" and argues that while averages have earned their place, they obscure the reasons behind voter decisions.
Cohen points to the tight Senate race in Maine as a prime example. Polls show Republican Sen. Susan Collins and Democrat Troy Jackson in a dead heat, with Collins relying on a significant share of voters who disapprove of President Trump. A September New York Times/Portland Press Herald/Siena poll had Collins up 3 points, within the margin of error. Other surveys show the race tied or Jackson ahead. Crucially, about a third of Collins supporters disapprove of Trump, and only half say they'd prefer a senator who backs him. Yet, as Cohen notes, the published cross-tabs don't break down how these cross-pressured voters answered questions about Trump, effectiveness, or ideology. The coverage stops at the margin.
Cohen argues that the real battle after Election Day will be over mandates and voter priorities. "Who has a mandate, and for what?" he asks. "What kinds of candidates do voters want, and which positions?" He contends that the data to answer these questions exist in public polls, but journalists rarely dig deep enough.
As an illustration, Cohen examines public opinion on data centers. Multiple polls—Gallup, Fox News, Emerson, The Economist/YouGov—found national majorities of 63 to 71 percent opposed to a data center near their home. Quinnipiac found 74 percent opposition in Pennsylvania. Heatmap and Embold Research tracked opposition rising from 42 percent to 75 percent between August and December 2025. A candidate could claim a mandate to block such projects, but Cohen notes the common thread is concern about electricity costs. Pew found Americans who had heard of data centers were six times as likely to say they're bad for home energy costs as good. Six in 10 told YouGov a nearby data center would raise their bills. When Echelon Insights asked voters to assume data centers would cover their own grid costs so bills didn't rise, 54 percent said local officials should allow them, while 34 percent said ban them. "Opposition is a headline," Cohen writes. "Together, these polls suggest a different question for the winners: Would voters accept data centers if developers paid the costs?"
Cohen sees AI as a tool to synthesize these tables, but he warns of pitfalls. "AI can also misread a column or mix up two groups," he notes. For example, a CNN poll found that 30 percent of Collins voters disapproved of Trump, but 20 percent of Trump disapprovers supported Collins. Confusing the denominators would tell the wrong story.
Cohen's prescription is transparency and journalistic judgment. Pollsters should disclose full question wording and sample details, and newsrooms should demand that before publishing any number. When AI flags an interesting finding, reporters should be able to verify it and then go deeper—recontacting respondents or drawing fresh samples to hear from cross-pressured voters. "Let them explain, in their own words, why they vote the way they do and what could change their minds," he writes.
Cohen acknowledges his 2006 stance was too rigid. "Good averages have earned their place," he says. But he insists that averages and prediction markets price outcomes, not reasons, and that the next step is to dig beneath the surface. "When the tables are public and the tools are this good, stopping at the margin is a lazy choice."
The broader context includes ongoing debates over election integrity and polling blind spots that could surprise midterm outcomes. Meanwhile, tariffs have emerged as a key midterm issue as Congress moves to reclaim trade authority. And as AI tools become more common in journalism, security agencies are grappling with AI-discovered vulnerabilities that could affect election infrastructure.
