Disconfirmation bias

The two traffic reports

A town is debating a new bike lane. A resident who opposes it reads two consultants’ reports that used the same method, a one-week count of cars before and after a trial closure. The one finding worse congestion strikes them as solid. The one finding no change, they point out, “only counted for a week, in summer, when half the town is away.”

The objection is fair, and it applies equally to the report they accepted. Both reports had the same week and the same season. Only the unwelcome one was checked for that weakness, so the difference in how convincing they seem comes from the conclusions, not the methods.

The rival camp’s study

On a running forum, a longtime advocate of minimalist shoes sees a post from a cushioned-shoe advocate citing a study on injury rates. She asks for the sample size, who funded it and whether it was peer-reviewed. A week earlier she shared a fellow minimalist’s post about a smaller study without checking any of those things.

Here the trigger is the source’s side rather than the argument’s content. The questions are good ones, and the post from her own camp needed them more. Many of the studies in this area manipulate exactly this, attaching identical information to one side or the other, and find the same asymmetry.

The careful reviewer

A researcher reviewing a paper that contradicts her own published findings reads its methods closely, finds that it relied on a convenience sample, and recommends rejection. For a paper confirming her findings the same month, she skimmed the methods, which also relied on a convenience sample, and recommended acceptance.

This is the respectable form: expert, effortful, and correct about the weakness it found. It is still the error, because the effort went only one way. Both papers had the same weakness, and only one was treated as disqualified by it.

• Biased assimilation. The outcome rather than the process: rating evidence that supports your view as more convincing and better done than matched evidence against it. Charles Lord, Lee Ross and Mark Lepper’s 1979 study named it.
• Selective search (confirmation bias in the narrow sense). Choosing which arguments to read in the first place, rather than how hard to check them. Charles Taber and Milton Lodge treated it as a separate tendency from disconfirmation bias, and grouped both under motivated skepticism.
• Myside bias. Keith Stanovich’s broader term for evaluating, generating and testing evidence in a way biased toward your own opinions. Disconfirmation bias is the evaluating part.
• Attitude polarization. The proposed consequence: two people with opposite views read the same mixed evidence and both come away more extreme. It has been much harder to show than the asymmetry itself (see Evidence).

• When the claim conflicts with strong independent evidence. A claim that contradicts a large, well-tested body of findings is more likely to contain a mistake, so it is reasonable to look harder. This is checking based on the evidence, not on which answer you’d prefer.
• When the source has a record of unreliability on this kind of claim, and you would discount a source on your side with the same record.
• When the standard is the same and the results differ. If one side’s arguments fail a checklist you apply to everyone, that is not a double standard.
• When the stakes justify extra checking of one claim, such as one you’re about to act on, and you’d check a welcome claim with the same stakes as hard.

The test: would you have raised the same objection if the argument had reached the conclusion you prefer, or come from your side?

The perpetual motion video

A physics student watches a video claiming a magnet-powered wheel produces more energy than it uses. She looks for the hidden battery, the edited cut and the measurement error. When her textbook says energy is conserved in a closed system, she doesn’t go looking for flaws in it.

The scrutiny is unequal, but it tracks the evidence rather than her preference. Conservation of energy is backed by an enormous number of independent observations, so a video contradicting it is far more likely to contain an error. She would check a surprising result she liked just as hard. Kari Edwards and Edward Smith, who named the bias, made this point themselves: searching harder for a flaw where a flaw is more likely can be rational.

The same checklist

A newspaper’s fact-checking desk applies one checklist to every candidate’s claims in a local election: is there a source, does the source say what was claimed, and is the figure current. One candidate’s claims fail far more often than the other’s, and the desk’s ratings reflect that.

The outcomes are unequal, but the standard isn’t. Every claim went through the same checks before anyone knew whose claim would fail. That makes this the third condition above. Disconfirmation bias is about uneven effort, not uneven results.

Edwards and Smith’s account is that an argument you disagree with sets off a deliberate search of memory for reasons it’s wrong. Because what you remember about an issue mostly fits your own view, the search finds counterarguments, and the argument ends up rated weaker. An argument you agree with sets off no search, so it is accepted as it stands. In their experiments, participants spent longer on arguments they disagreed with and listed more thoughts against them.

From the inside this feels like rigor, because every objection you find is a real one. That is Motivated reasoning‘s “illusion of objectivity”: you reach the welcome conclusion only because you can justify it. It also overlaps with Confirmation bias, since what you already believe decides which arguments get the hard check. In Taber and Lodge’s studies, people who knew more about politics showed the bias more, not less, because they had more counterarguments ready to use. And because the bias is easier to see in others (see Bias blind spot), the person showing it is usually the last to notice.

• Run the symmetry test before judging. Imagine the same study with the opposite result, or the same argument from your own side, and ask whether your objection would still occur to you.
• Fix the standard in advance. Decide what would count as good evidence before you know which way it points, and then apply that.
• Consider the opposite. In two experiments by Lord, Lepper and Elizabeth Preston, asking people to consider how they would have judged the evidence had it come out the other way reduced biased evaluation more than asking them to be fair and unbiased. That is one set of experiments, not a proven remedy.
• In someone else, pointing out the asymmetry (“you didn’t ask that of the other study”) is more productive than defending the argument they’re attacking, since their individual objections may well be right.

Status: contested. The tendency to rate the same information as better when it supports your side is well documented, and a 2019 meta-analysis found it across dozens of studies. But the evidence falls short of robust by this site’s criteria, and the interpretation is seriously disputed. No preregistered or multi-lab replication of the scrutiny process was located, and the meta-analysis tested for publication bias without correcting for it. Several researchers argue that some or all of the asymmetry may be a reasonable use of prior beliefs rather than a bias. Where the evidence falls between two labels, the less confident one applies.

• Lord, Ross and Lepper (1979) gave 48 students, half for and half against capital punishment, descriptions of two studies on whether it deters murder, one supporting each side, with methods varied so that each design appeared on both sides. Each group rated the study supporting its own view as more convincing and better conducted. Both groups also said their views had become more extreme, but that was their own report, not a before-and-after measurement.
• Edwards and Smith (1996) had 54 University of Michigan undergraduates read arguments on seven issues, including the death penalty and gay and lesbian adoption, one argument per side for each issue. On every issue they spent more time on the argument that conflicted with their prior belief, significantly so on five of the seven. They generated more counterarguments against the ones they disagreed with, and rated them weaker. The same paper argued that the extra scrutiny can be rational, since an argument you think is wrong is more likely to contain a flaw. But it also found steps that weren’t rational, such as people repeating the same counterargument as though each repetition counted again.
• Taber and Lodge (2006) ran two studies with Stony Brook students (126 and 136) on affirmative action and gun control. Arguments for participants’ own side were rated stronger, except among the least knowledgeable and those with weak views. Averaged across everyone, the extra reading time for opposing arguments was small, about 1–2 seconds, and only marginally significant. Among the most politically knowledgeable and those with the strongest views it was 4–7 seconds (25–50% longer). In thought listings, collected in the second study only, opposing arguments drew more thoughts, almost all critical. Polarization appeared only among the most knowledgeable participants, those with the strongest prior views, and those who processed the arguments most one-sidedly. The authors noted that earlier attempts to find polarization with before-and-after measures had mostly failed.
• Ditto and colleagues (2019) meta-analyzed 51 experiments with 18,815 U.S. participants in which matched information was made to support or oppose participants’ politics, by changing its content or its source. The average bias was modest (r = .245) and very similar for liberals (.235) and conservatives (.255), and it appeared in 39 of the 51 samples. Published studies showed a larger effect (.266) than unpublished ones (.139), though both were above zero, and a funnel-plot test found no sign of publication bias. The meta-analysis measured ratings, not reading time or counterarguments. It left out Taber and Lodge’s studies because their arguments for each side weren’t matched closely enough.
• The criticism. In a reply, Jonathan Baron and John Jost argued, among other things, that the pattern the meta-analysis measured is not necessarily an irrational bias at all. Ben Tappin, Gordon Pennycook and David Rand argue more generally that common designs in this area can’t separate wanting a conclusion from already believing it, since giving less weight to evidence that conflicts with your beliefs can be rational. They don’t claim that motivation plays no role, only that these designs often can’t show it.
• Polarization. A 2025 study by Yamil Velez and Patrick Liu showed people arguments, written for each participant, against the positions they cared most about. It found polarization only when the arguments were contentious and vitriolic, not for milder ones. Like the Backfire effect, the strong version of the claim, that opposing evidence leaves people more extreme, should not be treated as established.

What remains uncertain. How much of the asymmetry is bias rather than reasonable weighing of what people already believe; whether the scrutiny process holds up in preregistered tests; and how far findings from laboratory arguments, mostly about U.S. politics and with student samples, carry over to everyday disagreements.