Continued Influence Effect: Why a Retracted Fact Keeps Doing Its Work

The effect is a gap problem. When you read a story you build a model of the event, with causes connected to consequences, and the paint and gas cylinders arrive as the cause of the explosions, the oily smoke and the toxic fumes. A retraction removes the cause and leaves the consequences standing. Asked to explain them, the mind reaches for the only explanation it has ever had, and the retracted detail does the work even though its retraction is also in memory. This is why, across the 1994 experiments and much later work, people who recall the correction perfectly well still reason as though it had not arrived.

The account predicts what should help, and it does. A correction that supplies an alternative cause fills the gap, and in Johnson and Seifert's work the alternative was the one manipulation that reliably reduced reliance on the misinformation. Ecker, Lewandowsky and Tang found the same in 2010 with a story about a minibus crash: a plain retraction reduced references to the misinformation only slightly, while a retraction paired with an alternative account cut them by 42 to 67 percent on the questions that carried the effect, and a detailed warning before reading, explaining that corrections often fail, cut them by 53 to 63 percent. Nothing eliminated the effect. The title of that paper is the field's summary: warnings reduce but do not eliminate.

Two further mechanisms are usually named. Repetition makes a claim familiar, and a familiar claim feels true even after it has been corrected, which is why a correction that repeats the myth prominently can leave it more familiar than before. Retrieval failure is the other: the misinformation and the correction are stored separately, and at the moment of reasoning the original is retrieved without its tag. The meta-analyses agree on the shape. Chan, Jones, Hall Jamieson and Albarracín found across 52 studies in 2017 that debunking works, that misinformation persists after it, and that persistence is stronger when people have generated reasons for the original claim. Walter and Tukachinsky's 2020 meta-analysis of 32 studies with 6,527 participants found that corrections do not entirely eliminate the effect, and that they do least when the misinformation came from a credible source, had been repeated, or was corrected after a delay.

The warehouse fire belongs to Wilkes and Leatherbarrow, who published "Editing episodic memory following the identification of error" in the Quarterly Journal of Experimental Psychology in 1988. Participants read a fictitious news account as a series of time-stamped messages. Early messages placed cans of oil paint and gas cylinders in a closet near where the fire began; a later message stated that the closet had been empty. Inferences based on the old information continued to be drawn, the authors reported, although participants clearly recalled that it had been corrected.

Hollyn Johnson and Colleen Seifert took up the paradigm in "Sources of the continued influence effect," published in the Journal of Experimental Psychology: Learning, Memory, and Cognition in 1994, and gave the effect its name. Their participants read the fire messages and then answered open-ended inference questions, such as what might have caused the explosions, along with fact questions and a direct question about the correction. In Experiments 1A and 1B they tested whether timing mattered, with the correction either immediately after the misinformation or five messages later, and found that it did not: both correction groups made as many inferences from the volatile materials as a group that was never corrected, and far more than a control group that never read about the materials at all. Later experiments in the paper showed that supplying an alternative explanation, such as evidence of arson, reduced the inferences, while simply restating the correction did not.

The paradigm has been replicated many times. Ecker, Lewandowsky and Tang reproduced the pattern in 2010 with a different story and 125 participants. Laurent, Kosinski and Rusinek replicated Experiment 1A directly in 2023 with French-speaking participants, 21 who read the misinformation and its retraction and 23 controls, and found the same result: the corrected group drew more inferences from the false detail than the group that had never read it.

The effect is most visible where an early explanation is withdrawn and nothing takes its place. The rows below are illustrations, not findings from the studies cited on this page.

The research on corrections points to a short list of moves, none of which fully removes the effect.

• Give an alternative. A correction that says what did happen fills the gap the retraction opens. This is the best-supported remedy, from Johnson and Seifert's experiments onward.
• Lead with the fact, not the myth. State what is true first, mention the false claim only as much as needed, and finish with the fact again.
• Warn in advance where you can. Ecker and colleagues found that a specific warning about how corrections fail reduced reliance on the misinformation, while a vague reminder that facts are not always checked helped less.
• Repeat the correction. One retraction competes with a claim that may have been read several times.
• When you are the reader, notice when you are explaining a consequence with a cause you know was withdrawn. The memory check in the demo is the question to ask yourself: what did the correction say, and does my explanation still use the thing it removed?

These are procedures for writing and reading corrections, not a cure. The meta-analyses show that debunking reduces false beliefs and that a residue remains, and the residue is the effect this page describes.

Continued influence is one of several ways a correction can fall short, and it is the common one.

• It is not the backfire effect, the claim that a correction increases belief in the false claim. Backfire in that strong sense is rare and has failed to replicate in large studies; continued influence is a correction that works less than completely, and it is robust.
• It is not the misinformation effect, in which a leading question or another witness's account alters what a person remembers seeing. That is a change in memory for an event; continued influence is reasoning that keeps using a fact the person knows to be false.
• It is not the illusory truth effect, in which repetition makes a claim feel true. Repetition is one of the mechanisms that feeds continued influence, but illusory truth needs no correction to exist.
• It is not confirmation bias, the tendency to seek and weigh evidence for what you already believe. Continued influence appears in people with no stake in the story at all, reading about a fire in a warehouse they have never heard of.