Did you know? Someone can learn that the evidence supporting an idea was fabricated, acknowledge the correction, and still feel that the original conclusion is probably true. Psychologists call this family of effects belief perseverance. It is not simply refusing to listen. An initial explanation can become a self-sustaining mental model, leaving a conclusion psychologically available even after the example that first supported it has lost credibility.
The practical answer is to revise more than the isolated fact. Identify the exact claim, remove or downgrade the discredited evidence, examine the causal story built around it, search for genuinely independent information and state what conclusion the remaining evidence supports. A correction that merely says ‘that example was false’ may leave the original explanation untouched. A responsible review also distinguishes an unsupported belief from a belief independently supported by other reliable evidence.
This article explains belief perseverance using the original 1980 Anderson, Lepper and Ross research, the earlier 1975 Ross, Lepper and Hubbard experiments, and a modern 2022 review of misinformation and correction. It includes fictional worked examples, a practical evidence-revision method, limits and a changed-context exercise. None of the illustrative people or percentages below represent observed eduKate students.
Contents
- What persists when the original evidence disappears?
- What the original psychology experiments actually examined
- How a plausible causal story becomes independent of its example
- Belief, impression, inference and action are separate outcomes
- Belief perseverance is not the same as familiarity or confirmation bias
- Worked example: an evidence ledger for forty claims
- A precise diagnostic: what was removed, and what still supports the claim?
- Correcting the source and replacing the broken explanation
- Educational use: moving from a wrong mathematical rule to an accurate one
- Organisational use: when a mistaken incident story becomes policy
- How to test whether a correction actually changed reasoning
- Six questions that interrupt a self-sustaining explanation
- Eight mistakes when trying to correct persistent beliefs
- Changed-context exercise: from a fabricated study to a project estimate
- Research limits and how to avoid overdiagnosing people
- Frequently asked questions
- Primary research and the next reader route
What persists when the original evidence disappears?
A belief is a proposition a person treats as likely or true. Evidence is information that can raise or lower the credibility of that proposition. An explanation is a model connecting events through possible causes. These three are related but not identical. If a piece of evidence is discredited, a careful thinker should reconsider how much support remains. Yet the explanation built while considering that evidence may still feel plausible.
Belief perseverance refers to cases where a judgement persists beyond what the remaining evidence warrants, especially after an initial basis has been discredited. The effect is not defined merely by having a strong opinion or revising slowly. It concerns the relation between the strength of a belief and the evidential grounds that remain after correction.
Imagine a tutor showing a fictional class two made-up case descriptions suggesting that people who take bigger risks always learn mathematics faster. Students may construct explanations: risk takers practise more, tolerate mistakes and progress quickly. If the tutor then explains that the cases were invented and contain no actual observations, a student might still find the causal story attractive. That attraction is understandable, but it is not evidence that the original claim is empirically established.
The important distinction is between possible and demonstrated. Many causal narratives are internally coherent. A story that could happen does not show how often it happens or whether other mechanisms explain the observations better. Correcting a belief requires rechecking both the evidence and the inferential path connecting evidence to conclusion.
The skill is useful in school, organisations, science, everyday news and technical operations. When a team learns that a project incident was misreported, it should not continue blaming the same cause solely because the initial explanation was compelling. It should reopen the causal investigation with the remaining genuine records.
What the original psychology experiments actually examined
In a 1980 paper, Craig A. Anderson, Mark R. Lepper and Lee Ross investigated perseverance of social theories using fictitious case studies about the relationship between risk-taking and success as a firefighter. Some participants generated explanations for the apparent relationship. The studies examined whether theories could survive when the original case evidence was later discredited.
The authors reported that people could continue to endorse a causal relationship even after the initial evidential basis was undermined. They discussed how generating a causal scenario could contribute to that persistence. The paper’s author-hosted abstract makes clear that the experimental setting and the explanatory process are central to the finding.
Earlier work by Ross, Lepper and Hubbard in 1975 studied false feedback about performance and the later persistence of self-impressions and social impressions after debriefing. The experiments did not show that all participants held every false belief forever. They showed a measurable tendency in the particular experimental settings, and they explored whether forms of debriefing affected the result.
The lesson for modern readers is not that people are incapable of changing their minds. Instead, a correction can fail to remove the cognitive products created while the mistaken information was treated as true. The stronger claim—that every explanation causes perseverance in every population—would go well beyond what the experiments justify.
Research designs matter. The historical experiments involved specific tasks, sample characteristics and controlled misinformation. Everyday disagreements can include identity, incentives, group loyalty and a mixture of true and false evidence. Using the experiments responsibly means identifying the mechanism they illuminate while remaining cautious about quantitative claims outside the studied settings.
How a plausible causal story becomes independent of its example
Suppose you read a dramatic story about an athlete who changed one habit and suddenly won a competition. You explain the improvement: perhaps the habit increased confidence, improved focus and reduced mistakes. Later you discover that the story’s key details were invented. The original observation no longer supports the conclusion, but your explanations remain cognitively accessible and can feel like reasons to continue endorsing it.
This is not a magical effect. A causal explanation links concepts into a coherent network. Once a person has rehearsed that network, retrieving the explanation can become easy even without retrieving the source’s discredited status. Ease of retrieval is not proof of a claim. The correct next step is to ask whether the explanation has independent evidence or whether it is only a hypothetical mechanism.
Explanations can also hide alternatives. If you originally believed the athlete improved because of confidence, you may pay less attention to coaching, changed competition standards or ordinary random variation. When the original case collapses, the explanation should not be automatically promoted into fact simply because the alternatives were never examined.
One repair is to write the conclusion and its support separately. For the fictional sports story, record ‘the claim that the habit improves performance’ on one line and ‘the case study was fabricated’ on another. Then ask what new observations would actually discriminate between the confidence explanation and rival accounts.
Another repair is to generate a plausible counterexplanation without treating it as automatically true. For example, an athlete may report confidence after winning rather than win because of confidence. The exercise interrupts one-sided causal storytelling. Whether it changes a person’s conviction depends on context; it is a reasoning practice, not a guaranteed psychological intervention.
Belief, impression, inference and action are separate outcomes
A person’s stated belief may change while their spontaneous inference remains influenced by the old narrative. Someone may say ‘the article was corrected’ and still expect the original outcome in a later scenario. Others may change their explicit belief but retain a habit or social reaction. Researchers distinguish endorsement, inference and behaviour for this reason.
An impression of a person is not identical to a belief about a statistical relationship. If an individual was falsely told that a colleague performed poorly, learning that the report was random may not instantly remove every impression. Similarly, a student who once heard a misleading explanation of their mathematics ability may need real practice evidence before adopting an updated self-assessment.
Do not misuse the concept to label anyone who disagrees with you. A person may have independent information you have not considered, or your correction itself may be incomplete. Belief perseverance is a possible explanation for some residual judgments, not a diagnosis that can be imposed on a person after one conversation.
An observer should ask whether the original evidence truly was discredited, what other evidence remains, how the claim is framed and what exactly the person now believes. A correction that invalidates one anecdote does not logically invalidate a conclusion supported by multiple unrelated sources. Good reasoning can preserve a belief for good reasons as well as preserve it unjustifiably.
Evaluation therefore needs an evidence audit, not merely a confidence rating. Record what was believed before, what was corrected, what remained independently supported and what conclusion followed. This approach is fairer to the person and more useful for teaching.
Belief perseverance is not the same as familiarity or confirmation bias
The illusory truth effect concerns how repetition and familiarity can increase perceived truth. Belief perseverance concerns persistence after an original evidential basis is discredited. A claim may be both familiar and supported by a memorable explanation, but the mechanisms are not identical.
Confirmation bias concerns the tendency, in some contexts, to seek or interpret information in ways that favour an existing belief. It can reinforce perseverance, yet one can initially construct a causal theory from misleading information without subsequently searching for confirming data. The moment when the evidence is withdrawn is what makes the perseverance problem distinctive.
The naive realism guide examines people’s tendency to treat their own interpretation as simply objective and dismiss disagreement. That is a different issue from the source of a causal theory. Recognising the distinctions helps teachers choose targeted checks rather than reciting a list of bias names.
The continued influence effect, discussed in the 2022 Nature Reviews Psychology review, concerns the lingering influence of misinformation on reasoning after correction. It overlaps with some practical consequences of belief perseverance, but research traditions and experimental tasks differ. Avoid treating the two terms as perfect synonyms whenever the distinction matters.
A careful comparison asks three questions: did mere repetition make the claim familiar; did an inferred causal explanation survive removal of evidence; and did later information processing favour the existing belief? The answers can differ. This diagnostic sequence helps locate a repair more precisely.
Worked example: an evidence ledger for forty claims
Here is a fictional project, with deliberately constructed counts rather than real participant data. A small research group starts with forty claims about study routines. It finds independent support for twelve, discovers direct contradictions for eight, and has insufficient reliable evidence for the remaining twenty. The categories sum to 12 + 8 + 20 = 40.
Supported claims represent 12/40 = 30%, contradicted claims represent 8/40 = 20%, and unresolved claims represent 20/40 = 50%. If the group simply labels the twelve supported items ‘correct’ and everything else ‘false’, it confuses contradiction with uncertainty. If it labels the twenty unresolved items ‘probably true’ because they have attractive explanations, it makes the opposite error.
Now imagine that six of the eight contradicted claims were supported only by a fabricated demonstration. Their originally compelling stories should no longer count as observational evidence. Two contradicted claims have other, independently observed data that must be interpreted carefully. Both groups require reassessment, but the appropriate conclusion may differ because the total surviving evidence differs.
Suppose the team creates a revised statement for each of the forty claims, with one of three statuses: supported to a declared strength, contradicted under a specified condition, or unresolved pending evidence. The value of that ledger is not that it forces all beliefs to disappear. It prevents a persuasive story from silently substituting for evidence after the original basis is removed.
If the team later acquires strong evidence supporting five of the twenty unresolved claims, the new totals become seventeen supported, eight contradicted and fifteen unresolved, provided none of the earlier statuses change. Check: 17 + 8 + 15 = 40. The evidence ledger changes because new information has been independently assessed, not because someone repeated an explanation more confidently.
The percentages in this fictional example are bookkeeping. They are not estimates of how common belief perseverance is, nor are they measures of intervention success. The worked lesson is how to separate status categories, denominators and causal narratives so that future corrections remain coherent.
A precise diagnostic: what was removed, and what still supports the claim?
Begin with the exact proposition rather than a vague topic. ‘Exercise helps learning’ contains many possible claims about type of exercise, timing, age group and measured outcome. If one viral story is false, ask which narrow proposition it supposedly supported. Broad conclusions need broader evidence than one illustration.
Then identify the discredited item. Was a photograph miscaptioned, a quotation invented, a test result mistakenly assigned, or an entire case fabricated? Different corrections have different scope. A wrong date may not change the central conclusion; a fabricated outcome might invalidate the only evidence. The correction should specify what the record no longer establishes.
Next inventory any independent sources. Two webpages repeating the same incorrect report are not two independent observations. A claimed expert opinion that traces back to the fabricated story is also derivative support. Draw the source lineage where possible and avoid double-counting.
Finally ask whether the causal explanation has testable implications. If the belief predicts that a particular intervention should change an outcome, what observations would support or contradict that prediction? A mechanism stated so vaguely that every possible result fits it is not a strong basis for confidence.
This diagnostic can be performed privately by a learner before any interpersonal debate. The aim is to evaluate one’s own claim first. Applying the standard only to other people’s mistakes may leave the most important source of overconfidence untouched.
Correcting the source and replacing the broken explanation
A bare retraction may be necessary but insufficient. If a reader has built a detailed explanation around a false premise, correction should state that the premise was false and, where justified by evidence, describe a better-supported account. The replacement should not be invented merely to fill a story-shaped gap. It should be appropriately limited to what reliable evidence establishes.
Consider a fictional school report stating that a science experiment failed because students ignored instructions. Later, a timestamped laboratory record shows the equipment was calibrated incorrectly before the class began. A responsible correction withdraws the unsupported accusation and explains the genuine equipment finding. It does not simply say ‘the previous claim was inaccurate’ while continuing to blame the students in its recommendations.
A replacement explanation can be partial. Sometimes all that can be established is that the cause remains unknown. Stating uncertainty is better than choosing an attractive new villain. A sound mental model leaves room for unresolved causes and for future evidence that could refine the conclusion.
Specificity makes the correction auditable. ‘This photograph was from a different event and cannot establish attendance at the meeting’ is more useful than ‘misinformation was posted’. Include provenance, what changed and what remains unknown. Avoid needlessly repeating a false statement in a way that adds exposure without helping readers understand the correction.
The 2022 review discusses both preemptive and corrective interventions and the limits of knowledge revision. It does not license a promise that one format will reliably change every person’s view. The appropriate practical goal is a clear, evidenced route for revision that can be checked later.
Educational use: moving from a wrong mathematical rule to an accurate one
Belief perseverance has a useful analogue in learning, though one should not equate every student misconception with a controlled psychological experiment. A student may learn a mistaken rule, such as ‘multiplying always makes numbers larger’, because many early examples use positive whole numbers greater than one. The explanation feels coherent because those examples were real within a restricted range.
Then a teacher shows that 8 × 0.5 = 4, which is smaller than 8. The earlier general statement is false, but a student may still instinctively expect multiplication to enlarge a quantity. Simply marking the rule wrong may leave the underlying story unchanged. The teacher should rebuild the explanation around multiplication as scaling and explain the cases where the scale factor is below, equal to or above one.
Another example is algebraic cancellation. A learner who thinks the same digit or symbol can be crossed out anywhere may persist even after seeing a counterexample. The repair is to state the valid operation—division of an entire numerator and denominator by the same nonzero factor—and distinguish factors from terms. New examples must test whether the revised concept transfers to unfamiliar forms.
Good instruction pairs contradiction with a replacement model and independent practice. After explaining 8 × 0.5, ask the student to predict 12 × 0.25 and 12 × 1.5, then explain why the answers are 3 and 18. The changed numerical setting reveals whether the student still relies on the old rule or understands the new mechanism.
A tutor should not infer a psychological diagnosis from one wrong answer. Confusion, incomplete teaching, attention and limited vocabulary can produce similar visible responses. The pedagogical method is to inspect the explanation a student is using and supply evidence and practice suited to the actual misconception.
Organisational use: when a mistaken incident story becomes policy
Teams often need explanations quickly after a failure. An early report may attribute an outage to a recent deployment because the timing looks suspicious. People then generate a plausible chain: the new release changed a cache, the cache delayed requests and users encountered errors. Later logs may show that the actual cause was a separate network dependency. The original release theory can remain influential because meetings and mitigation plans already refer to it.
An effective review separates initial hypotheses from verified findings. Preserve the timeline of what people believed at each stage rather than retroactively rewriting the incident to make the first guess look accurate. Label which logs are primary, which are interpretations and which were unavailable during the first response.
If the original causal story is withdrawn, revisit decisions built upon it. Did the team disable a feature unnecessarily, blame the wrong owner or overlook the actual risk? A correction is operationally incomplete if the written conclusion changes but the resulting policy and alerts still implement the discredited theory.
The process should protect people who report that a previously popular explanation is wrong. Otherwise social incentives can make a narrative sticky for reasons beyond memory. Evidence review and decision rights matter as much as individual cognitive habits.
This connects to institutional forgetting, which concerns losing relevant knowledge. Belief perseverance concerns keeping a conclusion too long after its basis collapses. An organisation can suffer either failure, or both in different parts of its evidence system.
How to test whether a correction actually changed reasoning
Do not rely only on whether someone politely says ‘I understand’. A correction may change a person’s stated belief but not the predictions they make in a new case. A fair assessment can therefore examine the person’s explanation, confidence and response to a relevant transfer problem, while respecting privacy and avoiding manipulative testing.
Consider an invented classroom exercise with thirty learners. After a careful correction, twenty-four can accurately identify the false source, but only eighteen apply the corrected rule on a new example. Source recognition is 24/30 = 80%; transfer success is 18/30 = 60%. The twenty-percentage-point gap suggests that recognising the correction and using it are different tasks. It does not prove the cause is belief perseverance.
The remaining errors might arise from weak mathematics skills, ambiguous wording or fatigue. Ask the learners to explain their reasoning and compare it with the corrected model before selecting another intervention. A diagnostic that merely records ‘the learner resisted correction’ would overinterpret the limited observations.
To evaluate a correction programme systematically, compare relevant groups or repeated tasks with a sound design and meaningful baseline. Document the interval between correction and follow-up, task difficulty, prior knowledge and the criteria for a correct response. Statistical claims need actual observations rather than the fictional counts used for teaching here.
The best evidence of learning is often flexible transfer: using an updated rule when surface details change. This allows tutors to distinguish memorising ‘that example was fake’ from recognising how evidence and inference should change in a fresh situation.
Six questions that interrupt a self-sustaining explanation
Question one: What precisely was the original claim? Keep the proposition narrow enough to be tested. Question two: Which observation first made it persuasive? Trace the claim to its source rather than listing many derivative retellings.
Question three: What did the correction actually invalidate? A false photograph, an inaccurate quotation and a discredited causal study have different implications. Question four: Which independent evidence remains? Do not count a rehearsed explanation as if it were a new observation.
Question five: What alternative explanations predict the same pattern? A coherent story is not unique just because it is vivid. Question six: What future result would change my mind? If no possible observation could influence the conclusion, the claim may have been shielded from ordinary evidential testing.
These questions are tools for self-audit rather than interrogation. Asking them aggressively to embarrass another person may make constructive revision less likely. In a classroom, a teacher can model the questions using their own mistaken inference before asking students to apply them.
Importantly, the goal is not to force immediate reversal. A person with real independent evidence may rationally maintain a belief even after one source is corrected. The questions reveal whether maintenance is grounded in adequate support and whether the remaining uncertainty is honestly represented.
Eight mistakes when trying to correct persistent beliefs
Mistake one is confusing a retraction with a complete explanation. Removing a false example does not automatically repair the inferential model built around it. Mistake two is treating every persistent view as irrational when independent evidence may exist. Repair both by keeping claim, source and inference separate.
Mistake three is substituting a new unsupported story for the old one. An explanation should not be invented simply because uncertainty is uncomfortable. Mistake four is repeating the misleading narrative at unnecessary length while providing a tiny correction. Keep the accurate information clear and proportionate.
Mistake five is relying on confidence changes alone. A person may claim lower certainty while continuing the same invalid inference. Mistake six is ignoring source lineage: ten posts copying one mistake are not ten observations. Count genuinely independent evidence.
Mistake seven is shaming the learner or colleague, which can turn an evidence question into a social contest. Mistake eight is failing to update decisions that were made under the old belief. An operational correction should flow through relevant plans, teaching materials and policies.
Each repair can be checked. Can the person identify the withdrawn premise? Can they explain which evidence remains? Can they solve an unfamiliar case? Did the organisation update the procedure that relied on the wrong conclusion? These indicators are more meaningful than the length or confidence of a correction speech.
Changed-context exercise: from a fabricated study to a project estimate
Transfer the method to project planning. A committee once approved a system change after receiving a case study showing a 40% reduction in processing time. Later the case study is withdrawn because its numbers were fabricated. Several committee members still expect a 40% improvement because they have built a plausible explanation around the new software’s simplicity. What should the committee do next?
First withdraw the 40% figure as an evidential claim. It had no reliable observational basis. The committee may still have a hypothesis that software improvements could reduce processing time, but that is not the same as a measured benefit. Identify what the case study supposedly counted: perhaps total elapsed processing time, manual work time or error correction. The denominator matters.
Suppose a new controlled internal pilot, which is itself fictional for this exercise, records 50 tasks under the old process taking an average of 20 minutes and 50 comparable tasks under the new process taking an average of 18 minutes. The relative reduction from the old baseline is (20 − 18)/20 = 0.10, or 10%. The absolute reduction is 2 minutes per task, not 40%. These arithmetic facts depend on assuming the task groups are comparable; the pilot’s design and variability also need scrutiny.
The committee should update the forecast to reflect the genuine pilot evidence and its uncertainty, not protect the original 40% narrative by inventing special circumstances. It may still choose the new software for reliability, accessibility or other documented reasons, but those reasons must have their own evidence.
Change one condition: the pilot measured simple tasks while most deployment work is complex. The observed 10% reduction may not generalise. The proper next step is to test a representative sample or explicitly bound the prediction. This is the core transfer skill: invalidate a source, preserve only independently supported claims and update the decision without automatically choosing the opposite extreme.
Research limits and how to avoid overdiagnosing people
The original studies show that some judgments persisted after particular kinds of experimental evidence were discredited. They do not establish a fixed persistence rate for all ages, societies or beliefs, and they do not authorise labelling a person as biased from one disagreement. Behaviour in a controlled debriefing may differ from choices involving money, identity or long-held experience.
Some disagreements persist because the correction is not trusted, the evidence is genuinely ambiguous or the claim is supported by independent sources. Those are different explanations from belief perseverance. A strong debunker must hold their own correction to the same standards of provenance and causal inference that they demand from the original speaker.
Misinformation research investigates several overlapping mechanisms, including familiarity, motivated reasoning, correction design and continued influence. The 2022 review describes a broad evidence base, but no single tactic can guarantee revision. A responsible intervention should be evaluated on its own task and population.
A student mistake also need not be a psychological perseverance phenomenon. Missing prerequisites, confusing vocabulary, poor examples and insufficient practice can all produce repeated error. Teaching should begin with diagnosis and rebuild from the earliest unstable concept rather than treating learners as unwilling to change.
Finally, all numbers and invented examples in this article are for transparent reasoning. They are not estimates of prevalence, real intervention success rates or verified events involving the named educational ecosystem.
Frequently asked questions
What is belief perseverance? It is persistence of a belief or impression after the evidence that originally supported it has been discredited, beyond what the remaining evidence warrants.
Is it the same as lying? No. A person can sincerely retain an explanation that feels plausible even after learning the original example was invalid. Intentional deception is a separate matter.
Why can an explanation survive a correction? Generating a causal story may create links that feel persuasive independently of the original evidence. The original experiments examined how explanation was related to perseverance in specific settings.
Does every corrected belief need to be abandoned? No. A conclusion may still be justified by other independent, reliable evidence. The question is what support survives, not whether the original source was the only imaginable support.
How is this different from the illusory truth effect? The illusory truth effect concerns repetition and familiarity. Belief perseverance concerns continued endorsement after an evidential foundation is removed.
Can students experience something similar? Students can continue using a faulty explanation after a counterexample. That is a useful teaching analogy, but it does not prove an individual learner exhibits the psychological effect studied in laboratory experiments.
What helps repair it? State the correction precisely, audit remaining evidence, examine the causal story, consider alternatives and test the revised understanding in a changed-context task. No step guarantees immediate change.
Should teachers keep repeating the false idea so pupils remember the correction? Not as a blanket strategy. Communicate the accurate mechanism clearly, refer to the misconception only where it helps distinguish the valid rule, and check application rather than familiarity.
What if the evidence remains uncertain? Mark the claim unresolved and specify what further observation could discriminate among explanations. Uncertainty is often the accurate conclusion.
Primary research and the next reader route
Original theory-perseverance research: Anderson, Lepper and Ross (1980), Perseverance of Social Theories: The Role of Explanation in the Persistence of Discredited Information, Journal of Personality and Social Psychology 39, pages 1037–1049. The source is the author’s publication record and abstract.
Earlier experimental source: Ross, Lepper and Hubbard (1975), Perseverance in Self-Perception and Social Perception: Biased Attributional Processes in the Debriefing Paradigm, Journal of Personality and Social Psychology 32, pages 880–892.
Review source: Ecker and colleagues (2022), The Psychological Drivers of Misinformation Belief and Its Resistance to Correction, Nature Reviews Psychology 1, pages 13–29. The review distinguishes processes and interventions across multiple lines of research, with important limits on universal claims.
Related guides: illusory truth and familiarity, naive realism and disagreement, and pluralistic ignorance. These are distinct reader problems within the wider intelligence series.
If you are a learner, practise identifying a claim, its original evidence and its explanation as three separate items. If you are a teacher, use one changed-context problem to confirm whether the corrected mechanism transfers. If you lead an organisation, update decisions and documents that relied on withdrawn evidence.
Continue with How Intelligence Works and education that transfers into life. The core habit is simple: when evidence changes, revisit the reasoning built upon it rather than keeping only the most memorable story.