Backfire effect
Correcting a false claim was said to make the people most attached to it believe it more strongly, and the large replication attempts have not found that.
Correcting a false claim was said to make the people most attached to it believe it more strongly, and the large replication attempts have not found that.
A backfire effect is a reversal, not a failure. The correction has to leave belief in the misconception higher than it was before, or higher than in a group given no correction. Weaker outcomes have their own names and are not the same thing. Under the continued influence effect, entry 192, people go on using discredited information while still moving in the direction the correction pushed.
Under belief regression they update and then drift back. Three named mechanisms have been proposed for a true reversal. The worldview version says a correction that threatens someone's political identity provokes counterarguing. The familiarity version says repeating a myth in order to deny it makes it feel truer. The overkill version says piling on counterarguments is less persuasive than offering a few.
Swire-Thompson, DeGutis and Lazer review the first two and treat them as distinct psychological claims. The overkill version comes from the 2011 Debunking Handbook rather than from that review, and the 2020 rewrite reports no evidence for it.
Failed to replicate Wood and Porter tested 52 issues on more than 10,100 subjects and found no correction that triggered a backfire.
Wood and Porter ran five experiments on issues chosen as the polarised ground where a reversal should be easiest to produce. Their samples were 3,127, 2,801, 977 and 1,333 respondents from Mechanical Turk, plus 995 from the Lucid national panel and 1,024 from Turk fielded together on 6 to 8 June 2017. Across all of it they report no correction capable of triggering backfire.
Corrections improved the average respondent's accuracy on 85% of issues among liberals, 96% among moderates and 83% among conservatives. They also re-ran Nyhan and Reifler's own Iraq article and randomised the survey item. With the original long item, respondents ignored the correction and inferred nothing from it, which the authors call a partial replication. With a plain item reading that US forces did not find weapons of mass destruction after the 2003 invasion, liberals took the correction on board.
A further experiment wrote simple, moderate and complex versions of items on six issues, and found no backfire in any of the 18 combinations. Conservatives did show less factual uptake as items grew more elaborate. An earlier working paper from the same project, dated 5 August 2016, covered four experiments, more than 8,100 subjects and 36 issues, and reported one issue capable of backfiring, which was Iraq.
Haglin re-ran Nyhan and Reifler's 2015 flu vaccine study inside a larger study of flu vaccines and did not reproduce the effect. Swire-Thompson and colleagues then supplied a mechanism for the pattern of failures. In two near-identical longitudinal experiments, analysing 388 of 423 and then 532 of 600 participants who rated 21 misinformation items, no item backfired more under correction than under a test-retest control.
How often an item appeared to backfire tracked how unreliable the item was, at rho of -0.61 and -0.73. Item importance and strength of belief, the worldview markers, predicted nothing. Nyhan wrote in 2021 that backfire effects are extremely rare in practice, and that his own results had been misread as showing that all corrections are counterproductive. The Debunking Handbook was rewritten in 2020 and now tells communicators not to withhold a correction for fear it will backfire.
The corrected rumour at a residents' meetingillustration
A council officer presents figures refuting a rumour about a new development. Afterwards the residents most opposed to the plan believe the rumour more firmly than they did before. Belief moving up, rather than merely failing to move, is what the name claims.
The leaflet that hardens its readerillustration
A health leaflet sets out the evidence against a myth about a vaccine. Under the backfire claim the parents most hostile to that vaccine come away more convinced of the myth than parents shown no leaflet at all.
The fact-check a supporter readsillustration
A supporter reads a correction of a statement by the candidate he backs, and rates the false statement as more accurate than he did beforehand. Holding to it would not count; the correction has to push belief upward.
Cases marked as illustrations describe nobody in particular. They are there to make the pattern recognisable, not to report that a particular thing happened.
The four published experiments used a five-point agreement scale. Each looked for an interaction between the correction and a seven-point ideology scale. Study 1, run in autumn 2005 with 130 participants, corrected the claim that Iraq had weapons of mass destruction before the 2003 invasion. The correction by ideology interaction was 0.359 with a standard error of 0.127, and the marginal effect of the correction was positive and significant for subjects right of centre.
In the raw data, agreement among conservatives went from 32% in the control condition to 64% with the correction, against 22% to 13% for everyone else. Study 2, run in spring 2006, fielded three experiments on 197 respondents, of whom 195 entered the models. The tax cut experiment produced the second reversal, an interaction of 0.198 with a standard error of 0.089, and agreement among conservatives rising from 36% to 67% against 31% to 28% for the rest.
The second Iraq experiment went the other way. Its interaction was -0.258 with a standard error of 0.097, meaning the correction worked on conservatives. Nyhan and Reifler then ran what they call a post hoc model on the 34 people who named Iraq the most important problem facing the country, eight of them right of centre. The three-way interaction was significant, but the marginal effect for the most committed conservatives was null, and in their own words just missed significance at the 0.10 level.
The stem cell experiment found no reversal, only a correction that worked on conservatives and moderates and not on liberals. Two of the four backfired, one corrected conservatives as intended, and one did neither.
5 primary research5 replication and reanalysis7 commentary and critique
Their paper "When corrections fail" reports four experiments in which undergraduates read a mock news article carrying a politician's dubious claim, with or without a paragraph correcting it. The paper coined the label, quoting it on first use and dropping the quotation marks thereafter. Nyhan wrote in 2021 that he and Reifler ran five experiments in 2005 and 2006 with undergraduates at a Catholic university in the Midwest.
The claim, and the replication this entry reads its figures from.
What failed is the group-level claim. Wood and Porter say plainly that their result does not make backfire categorically impossible, only that they could not locate a single instance of it. Isolated subgroup findings were published and have not all been re-tested. Nyhan, Reifler, Richey and Freed's trial of 1,759 US parents found that refuting the MMR and autism link reduced the misperception, yet lowered vaccination intent among parents holding the least favourable vaccine attitudes.
Wood and Porter exclude that result from their own definition, because the reversal sits on intended behaviour rather than on the corrected belief. Swire-Thompson and colleagues note that worldview backfire effects have almost exclusively been found in political or attitudinal subgroups rather than in whole samples. They report that 81% of the backfire effects in their review came from single-item measures.
The share found with single items, 37%, exceeded the share found with multi-item measures, 8%, at Z = 2.96, p = .003. They also leave the door open in principle, saying a familiarity boost could in theory outweigh a correction, for instance under divided attention, and that this has not been observed. Wood and Porter used Mechanical Turk and an online panel, while the 2010 studies used undergraduates for course credit.
No re-run of the 2010 protocol on its original population was located. A search on 21 September 2026 of the Psychological Science Accelerator's project list found nothing on correction or debunking, and no multi-laboratory attempt on the effect was located.
Each of these is named in this entry's own prose as something it gets taken for. Where the two came out differently, that is worth knowing before you quote either.
The backfire effect is not the continued influence effect, entry 192, and the second one replicates. Entry 192 carries the pooled evidence for it. One figure from there is worth repeating, because it is the one people quote against corrections. Chan, Jones, Hall Jamieson and Albarracin pooled eight research reports, 20 experiments and 52 statistically independent samples, N = 6,878.
That is not 52 studies. Their persistence figure is the gap between a debunked group and a control shown neither the misinformation nor the correction. The d ranges in their abstract span estimation methods rather than a confidence interval. Persistence runs from 0.09 under fixed effects, with a 95% interval of 0.07 to 0.12, to 0.97 under random effects.
The authors flag the fixed-effects model of persistence as the one case that is not large, and heterogeneity across the analyses is close to 99%. Corrections work and leave a residue whose size depends heavily on how it is estimated. That is the finding people reach for when they say facts do not change minds, and it says the opposite.
Nor is the familiarity version the illusory truth effect, entry 27. Repetition without a correction does raise belief, and repetition inside a clear correction does not. Swire-Thompson and colleagues report a meta-analysis of 10 studies finding no difference in belief updating whether or not the myth was repeated in the correction, and credit it to Walter and Tukachinsky.
Reactance, entry 86, is a third thing again, a response to a threatened freedom rather than to a fact. Nothing here was retracted and none of this is a scandal. Nyhan and Reifler labelled one analysis post hoc in 2010 and never called the others exploratory. They later collaborated with Wood and Porter, and Nyhan set out his revised reading in PNAS. His argument there is that misperceptions persist for other reasons, including corrections whose effects decay, poor targeting, and a steady supply of contrary cues from political elites.
Every claim on this page was held against these sources on 2026-09-21. Nothing here is written from memory. Written and checked by Krishna Chagti; how entries are written and corrected. Found a mistake? Report it, with the sentence and a source.