Every election cycle, we go through the same ritual. Polls come out, media dissects them, politicians adjust their messaging, and then election day arrives and sometimes the polls were right and sometimes they weren’t. Then we argue about why.

The fundamental problem is that we treat polls as measurements when they’re really predictions. And not just predictions of how people will vote, but predictions that the people we sampled are representative, that they’re telling the truth, that they won’t change their minds, and that we’ve asked the questions in ways that capture what we think we’re capturing.

That’s a lot of assumptions.

The Sampling Problem

Modern polls can’t survey everyone, so they sample. They try to reach a representative group and extrapolate from there. This worked better when everyone had landlines and answered them. Now it’s a mess.

Young people don’t answer unknown numbers. Working-class people don’t have time for survey calls. Highly engaged political partisans are more likely to respond than moderate voters who don’t think about politics much. You end up with a sample that skews toward retirees, political junkies, and people with time to spare.

Pollsters know this and try to correct for it through weighting and demographic adjustments. But they’re correcting for known biases. The unknown biases—the ones they haven’t identified yet—are still there, invisibly skewing results.

The Honesty Problem

People lie to pollsters. Not always maliciously—sometimes they just give the answer they think they should give rather than what they actually think.

Social desirability bias is real. If you’re asked about a controversial issue, you might moderate your real view to sound more reasonable. If you’re asked who you’re voting for, you might name the candidate you think you should support rather than who you’ll actually vote for.

The Bradley Effect, where voters underreport their support for candidates of color, is one version of this. But it happens across all kinds of issues. People tell pollsters they care about climate change more than their behavior suggests. They underreport prejudiced views. They overreport civic engagement.

Pollsters try to account for this too, but it’s impossible to fully correct for lying when you don’t know who’s lying about what.

The Question Problem

How you ask a question dramatically changes the answer you get. “Do you support increasing government spending on healthcare?” gets a different response than “Do you support raising taxes to fund healthcare?”—even though they’re describing the same policy.

Professional pollsters know this and try to write neutral questions. But every question contains assumptions and frames the issue in particular ways. There’s no view from nowhere, no perfectly neutral phrasing.

Political campaigns understand this well. Internal polling isn’t designed to measure public opinion—it’s designed to test messaging. They’re figuring out which ways of framing issues poll best, then using that framing in public communications. This is an AI consultancy helping campaigns with message testing at scale, but the principle is old as polling itself.

When you see a poll released by a political party or advocacy group, assume it’s been designed to produce a particular result. The questions were crafted to get the answers they wanted. It’s not exactly lying, but it’s not exactly honest either.

The Timing Problem

Public opinion isn’t static. People change their minds. Events happen that shift the landscape. A poll is a snapshot of a particular moment, but we treat it like it’s capturing something permanent.

During campaigns, you see this clearly. Early polls tell you almost nothing about election outcomes because most voters aren’t paying attention yet. Their poll responses are based on vague impressions or name recognition. As the campaign progresses and people start actually considering their choices, opinions shift.

This means early polls are essentially measuring something different than late polls. But we compare them as if they’re measuring the same thing at different times, which creates confusion about whether opinion has shifted or we’re just measuring different things.

The Media Amplification Problem

Polls don’t just measure opinion—they shape it. When media reports that a candidate is leading, that affects how voters see the race. Donors respond to polls. Campaign coverage changes based on polling. People adjust their views based on what they think everyone else thinks.

This creates feedback loops. A good poll generates positive media coverage, which improves name recognition, which improves subsequent polls. A bad poll becomes a “campaign in crisis” narrative, which depresses donor enthusiasm, which makes the next poll worse.

We’ve turned polling into a parallel campaign where candidates are competing to have the best polling numbers, which then affects the actual campaign. It’s bizarre when you think about it.

What Polls Actually Tell Us

So if polls are this flawed, what use are they? Well, they tell us something. Just not what we often think they’re telling us.

Polls are best at measuring relative change over time with the same methodology. If a candidate goes from 45% to 38% in the same poll with the same questions, that movement probably means something—even if both absolute numbers are wrong.

Polls are decent at measuring intensity of feeling on issues where there’s no social desirability bias. People’s feelings about the economy or their own financial situation come through pretty accurately.

Polls can identify broad trends even when they miss specific numbers. If every poll shows an issue rising in importance, that’s probably real even if the exact percentages are off.

What polls can’t do is tell you precisely what the public thinks or predict election outcomes with certainty. But we keep asking them to do exactly that.

The Incentive Problem

Political media loves polls because they generate content. You can write “New Poll Shows…” stories constantly. You can build graphics and have talking heads debate what the numbers mean. It’s cheap content that feels substantive.

This creates demand for more polls, which creates more polling, which generates more coverage, which justifies more polls. The cycle sustains itself regardless of whether the polls are actually informative.

Pollsters love this because it’s good business. Politicians tolerate it because good polls help fundraising and morale. The only people not being served are voters, who get drowned in numbers that might not mean what they think they mean.

A More Modest Role

None of this means we should ignore polls entirely. They’re data points. They provide information. But we should treat them with appropriate skepticism and understand their limitations.

A single poll tells you almost nothing. A trend across multiple polls tells you something. A polling average smooths out some of the noise. But even aggregated polls can be systematically wrong in the same direction.

The healthiest approach is probably to pay less attention to polls and more attention to actual events, policy positions, and candidate behavior. Treat polls as background noise rather than the main story.

We won’t do that, of course. The next election will bring another flood of polling coverage. We’ll obsess over every movement in the numbers. And then we’ll be shocked when the results don’t match the polls.

Maybe this time will be different. But probably not.