Belief perseverance

The quiz with random scores

Someone takes an online “sense of direction” quiz and is told they scored in the bottom 10%. That evening they think back to the times they got lost in a new city. A week later the site announces that a bug had assigned everyone random scores. When a friend asks them to navigate on a road trip, they say: “Better not. I’ve never been great with directions.”

The score was the only new evidence, and it has been withdrawn. But in the meantime they recalled memories that fit it and built a story about themselves, and that story outlived the score. This is the classic form of the effect: early experiments gave people false feedback about how well they did at a task, told them afterward that the feedback was fake, and found that their self-assessments still leaned the way the feedback had pushed them.

The corrected closure story

A local news site reports that a bakery closed after failing a health inspection. The next day it corrects the story: the inspection was fine, and the owner simply retired. A month later, when the owner opens a small café, a neighbor says: “I’d be a bit careful there. Wasn’t there something about hygiene?”

The neighbor may well remember that there was a correction, yet the original explanation of the closure still colors their judgment. In research on misinformation this is called the continued influence effect: a retracted detail keeps shaping inferences, especially when the retraction leaves a gap in the story that the false detail used to fill.

The plausible mechanism

A gardener reads that a study found coffee grounds doubled tomato yields, and decides it makes sense: grounds add nitrogen, and tomatoes like nitrogen. Later she learns the study was withdrawn because its data were made up. She says: “Well, it probably still helps a bit.”

This is the less obvious form. She did update, and “a bit” is weaker than “doubled”, so there’s no defiance here. But her remaining confidence rests on an explanation she thought of because of the study, not on any evidence that grounds help. Partial updating like this is what research usually finds, and it is still perseverance: the belief is stronger than it would be had she never read the study.

• When other evidence supports the belief. If a retracted study was one of several independent studies pointing the same way, the belief shouldn’t collapse with it.
• When you held the belief before. A discredited piece of evidence should return you to where you were before it, not to disbelief.
• When the discrediting is itself doubtful. A vague claim that a report “was debunked” is weaker than a clear retraction from its source, and it’s reasonable to give it less weight.
• When the evidence was weakened rather than eliminated. If a study turns out to be smaller or less careful than reported, a partial update is the right size.

The test: if you had never seen the discredited evidence, would you believe this as strongly as you do now?

Your own visits still count

A café’s online reviews included a detailed complaint that its prices had doubled. The review is later removed as fake. A regular who has noticed the prices climbing on her own receipts still thinks the café has gotten expensive.

Her belief doesn’t rest on the fake review. Her receipts are independent evidence, and they’re untouched by its removal. That’s the other evidence condition: what should go is only the extra confidence the review added.

A rumor of a retraction

Someone reads a well-known study on how long people remember phone numbers. A post in a forum claims “that study was retracted years ago”. The journal’s page shows no retraction or correction, so they keep treating the finding as sound.

Declining to drop a belief because of an unverified claim isn’t perseverance. The original evidence was never actually discredited, which is the discrediting is doubtful condition.

Researchers point to explanation. People tend to make sense of new information by working out why it would be true, and in the original experiments, beliefs persisted more when people had generated such explanations. The explanation doesn’t mention the evidence it came from, so when the evidence is discredited, the explanation is still there, still pointing the same way.

A related account from misinformation research is about gaps in a mental model. If a false detail explained an event (why the bakery closed), retracting it leaves the event unexplained, and the familiar explanation stays the easiest one to reach for. False information also remains in memory alongside the correction, and can come to mind without it. Once a belief has formed, later information tends to be read in its light, a relative of Confirmation bias, and first impressions exert a pull much like an anchor.

Several remedies have been tested:

• Replace the explanation, don’t just delete it. In continued-influence experiments, supplying a plausible alternative cause reduced reliance on the discredited one, and a meta-analysis found corrections worked better when they were coherent.
• Argue the other side. The original researchers proposed having people explain why the opposite could be true. A 2023 experiment found that this helped, and that a correction repeating that the evidence was invented, noting the opposite might be true and giving arguments for it, helped more; a 2024 follow-up found that combining it with a short explanation of the bias and a warning about it worked better than either alone.
• Discredit the source of the evidence, not only the result. In debriefing research, as summarized in a 2024 study, telling participants that the test itself was fake worked better than telling them only that their score was.
• Correct early and clearly. Corrections were less effective when misinformation had been repeated, came from a credible source, or went uncorrected for a while.

For yourself, run the test above. If your answer to “why do I believe this?” is an explanation rather than evidence, ask where the explanation came from.

Status: robust. The central claim is that people don’t fully update when a belief’s evidence is discredited. It has been found in debriefing experiments since 1975, in a large literature on corrected misinformation, in a meta-analysis whose persistence effect held up across several checks for publication bias, and in preregistered experiments. How much belief survives varies greatly, and some studies find none.

The debriefing paradigm.

• Ross, Lepper and Hubbard (1975), as described by the same researchers in 1980, gave participants false feedback on how well they could tell authentic suicide notes from fake ones, then thoroughly debriefed half of them about the feedback being fake. Although participants understood and accepted the debriefing, their ratings of their own ability and predictions of future success still followed the discredited feedback. Observers who had watched showed the same pattern in their impressions.
• Jennings, Lepper and Ross (1981) found the same with a more everyday task: people’s impressions of how persuasive they were, after trying to persuade another student to donate blood, persisted after they learned that the other student had followed a script. Asking people to explain their performance did not increase perseverance, contrary to the researchers’ prediction.
• Anderson, Lepper and Ross (1980) extended it from self-perceptions to beliefs about the world. Participants read two case studies suggesting that risk-taking made someone a better or a worse firefighter, and some were then told the cases were fictitious. In the first experiment, the beliefs “survived virtually intact”; the small drop after debriefing was not statistically significant. In the second, debriefing made beliefs less extreme but did not erase them. People whose explanations offered a general reason for the relationship tended to show more perseverance. This is the strong version of the effect; later work more often finds substantial but incomplete updating.
• Greitemeyer (2014) told participants about a scientific finding and then, for some, that the article had been retracted for fabricated data. The debriefed group believed the finding less than those not told, but more than people who had never read it. Analyses pointed to the causal arguments people had generated for the finding as what kept the belief going.
• A 2007 replication in the suicide-notes paradigm, as summarized by Miketta and Friese (2024), found that a debriefing explaining that the test itself was invalid removed the effect of the false feedback better than a standard debriefing. Miketta and Friese’s own experiments found that a debriefing of that kind did not undo the lowered mood caused by a staged experience of social exclusion; that concerns feelings rather than beliefs, but it points the same way.

Corrected misinformation. Research on the continued influence effect measures the same thing with news-like reports.

• Johnson and Seifert (1994) showed that people went on using a corrected detail when making inferences about an event, even when they remembered the correction, and that an alternative explanation reduced this. Lewandowsky and colleagues (2012) reviewed why retractions so often fail to fully undo misinformation and what makes corrections more effective.
• Chan and colleagues (2017) meta-analyzed 52 samples (6,878 participants). Debunking had a large effect, but belief in the misinformation after debunking remained well above the level of people never exposed to it (d = 0.75 to 1.06). Persistence was stronger when people had generated reasons supporting the misinformation. Checks for publication bias were mixed; selection models changed the persistence estimate only slightly.
• Walter and Tukachinsky (2020) meta-analyzed 32 studies (6,527 participants) and concluded that, on average, corrections do not entirely eliminate misinformation’s effect. The residual effect they report is small (r = .05 in magnitude), much smaller than Chan’s estimate; the two analyses pooled different studies and measures.
• Ecker, Butler and Hamby (2020), a registered report of three preregistered experiments (2,279 participants in total), tested narrative against plain corrections. With fictional event reports, corrections strongly reduced reliance on misinformation, but a small, reliable continued influence remained with both formats, both immediately and two days later.

Where it doesn’t appear. In Greitemeyer and Sagioglou’s (2015) two studies, participants who learned that a researcher accused of misconduct had been exonerated ended up rating the researcher about as favorably as people who had never heard the accusation. Mickelberg and colleagues (2025) found a continued influence of a retracted negative behavior on impressions of a person in one of three experiments; in the other two it was fully discounted. And in a 2024 experiment by Siebert and Siebert, 504 of 876 participants shifted their opinion after misinformation, and 364 were still shifted, by the study’s threshold, after it was retracted: common, but not universal.

Not the backfire effect. Belief perseverance means updating too little: the belief moves toward the correction but stops short. The Backfire effect was the claim that a correction makes a belief stronger, moving it the wrong way. Large tests have rarely found backfire, which is fully consistent with robust perseverance. The two are often blurred together in popular writing, and the failure of one says nothing against the other.

What remains uncertain. The size of the effect ranges from beliefs left nearly intact to complete updating, depending on the kind of belief, how the evidence was discredited, and how the belief is measured. Many experiments use fictional events and short delays, so less is known about beliefs held for years. The explanation-based account is well supported in the original work but has had fewer direct tests since.