Mandela Effect

​

What is the Mandela Effect?

The Mandela Effect, as counted here, is a shared and specific false memory: many people remember one wrong detail of a public icon, and that detail is not the canonical version.

The Mandela effect is an Internet phenomenon describing shared and consistent false memories for specific icons in popular culture.

That sentence is the abstract of Deepasri Prasad and Wilma A. Bainbridge, Psychological Science, volume 33, issue 12, December 2022, pages 1971–1988. The text read here is their accepted manuscript. The journal record lists electronic publication on 11 October 2022. The DOI is 10.1177/09567976221108944. Both authors are affiliated with the Department of Psychology at the University of Chicago. They use Visual Mandela Effect for the picture form. This page uses Mandela Effect for that shared error. It does not use the name for every fuzzy memory, and it does not use it for a claim that the record itself was altered. Illusory Truth Effect is a higher truth rating after a plausible statement is heard again. Mere Exposure Effect is a change in liking. Neither one is a room that picks the same missing feature.

​

The Mandela Effect in 3 Depths

• Beginner: Three pictures, one of them official. For some icons, people pick the same unofficial picture and feel sure. The shared wrong detail is the effect. A low score by itself is not.
• Practitioner: Write the feature before you quote a percent. A monocle, a cornucopia, a tail, a black tail-tip, a gold leg, and a missing gap are six different errors. They are not one rate.
• Advanced: The 2022 chi-square of independence treated one person’s single choice as if the options were separate people, and the statistic grew. A goodness-of-fit reanalysis keeps six icons in the original data. Waldo’s corrected p is 0.052. Do not move either result onto a lyric, a death date, or a slogan the test never showed.

​

Origin

The counted test is Prasad and Bainbridge, “The Visual Mandela Effect as Evidence for Shared and Specific False Memories Across People.” The manuscript says it was accepted on 28 March 2022. The review board was the University of Chicago, protocol IRB19-1395. Data and pictures are at osf.io/7cmwf. Experiment 1 was exploratory. It was not pre-registered. Their reference list dates a page by Fiona Broome to 9 September 2010 and titles it “The Mandela Effect — The Article That Started It All in 2010,” at mandelaeffect.com/nelson-mandela-died-in-prison/. That page did not open here, so this article does not quote it. David Wilton, on Wordorigins.org, 7 March 2025, says the earliest example of the phrase he could find is a post by Reece on the blog The Wood Between Worlds. The essay dates that post 23 June 2012. The source line dates it 23 June 2014. This page did not open the blog, so it does not choose between those two dates. Wilton writes that news articles credit Broome, a paranormal researcher, with coining the phrase and with a parallel-universe idea, and that he found no evidence she originated either one. Wilton quotes Reece: the name comes from a belief that Nelson Mandela died in the 1980s. Wilton says Mandela’s death was in 2013. He offers, as an untested guess, that some of those memories mix Mandela with Steve Biko, who died in police custody in 1977. No percent for the prison memory is on this page. It was not one of the 40 icons. Wilton’s earliest newspaper example is Russell Smith, The Globe and Mail (Toronto), 17 August 2015. He also says some readers treat the shared error as a glimpse of another universe. The icon experiments do not test that idea. A false memory does not require a second universe.

​

Key Points

Four readings keep a shared wrong detail from turning into a changed record, a failed drawing, or a percent that belongs to a different picture.

​

Applications

Use the effect as a reason to open the file. Do not use a laboratory percent as the error rate of a meeting, a lyric, or a family story.

​

Case Study

Prasad and Bainbridge wanted a test of the informal reports, not another list of them. They built 40 sets. Each set had the official picture and two edited ones. Edits were made in GIMP. Only one feature changed. Across 80 edits there were 26 removals, 21 additions, 24 changes, 3 turns, and 6 color changes. The edits were meant to be relatively consistent with the usual schema, and the type of edit did not point to the official picture. For the six reported icons, the reported false feature was one of the two edits. Each of 100 adults on Amazon Mechanical Turk, screened for the United States and for English, saw all 40 sets. A person who left more than half the questions blank was excluded. The 100 are the people who finished. Mean age was 39.4 years, standard deviation 12.3. Each image had 98.1 responses on average. Order of the sets, and order of the three pictures, was random. The question was which picture was the official one. Confidence, familiarity, and a rough count of past views came after the choice. They then labeled Manipulation 1 as whichever wrong picture had drawn more choices. That label is applied after the choices. For the six, the reported false feature had been built in advance as one of the two edits. The test asks whether the more popular wrong picture beat the official one. Seven icons went that way on the independence test the paper printed. The replication team reconstructed that test image by image. Those reconstructed lines are not printed in the 2022 paper. The paper prints only the summary, chi-square at least 6.089 and p below 0.014 for 39 icons. The team also says the N in that reconstruction is more than double the number of people, because the wrong test counts the options as if they were separate. On the goodness-of-fit test, original data, the lines are these. C-3PO, χ²(1, N = 91) = 33.24, p = 8.138 × 10⁻⁹. Curious George, χ²(1, N = 93) = 23.75, p = 1.095 × 10⁻⁶. Fruit of the Loom, χ²(1, N = 82) = 3.95, p = 0.047. Monopoly Man, χ²(1, N = 83) = 11.58, p = 6.673 × 10⁻⁴. Pikachu, χ²(1, N = 76) = 7.58, p = 0.006. Volkswagen, χ²(1, N = 91) = 16.71, p = 4.345 × 10⁻⁵. Waldo, χ²(1, N = 86) = 3.77, p = 0.052. Tom, from Tom and Jerry, χ²(1, N = 89) = 1.36, p = 0.244. Waldo and Tom are not under 0.05 in that column. The N on each line is the contrast, not the sample of 100. The same team collected a new sample of 389 and ran the same corrected test. The report page read here does not print the recruitment platform or the ages. C-3PO, χ²(1, N = 359) = 99.50, p = 1.960 × 10⁻²³. Curious George, N = 384, χ² = 70.04, p = 5.806 × 10⁻¹⁷. Fruit of the Loom, N = 369, χ² = 10.08, p = 0.001. Monopoly Man, N = 378, χ² = 4.67, p = 0.031. Pikachu, N = 304, χ² = 39.80, p = 2.810 × 10⁻¹⁰. Volkswagen, N = 362, χ² = 54.14, p = 1.864 × 10⁻¹³. Waldo, N = 351, χ² = 26.81, p = 2.249 × 10⁻⁷. Tom, N = 367, χ² = 23.57, p = 1.207 × 10⁻⁶. In that sample, eight icons meet the corrected rule, including Waldo and Tom. Neither analysis prints a correction for testing 40 icons. Fruit of the Loom at 0.047, and Monopoly at 0.031 in the new sample, sit near 0.05. The lesson is a specific wrong picture, on a named icon, under a named test. The cornucopia and the monocle are two of those pictures. They are not a measurement of how often a company, a classroom, or a family misremembers a fact the experiment never showed.

​

Boundaries and Failure Modes

The Mandela Effect is the wrong summary when a shared story has no counted feature, when the mistakes do not match, or when one p value is asked to cover every icon. The first boundary is the story the test did not contain. The prison memory is the name’s occasion, not a cell in the table. Wilton’s Biko suggestion is his suggestion. The spelling Berenstain is a popular example in his essay. Movie lines are not in the 40. Hindsight Bias is a past event feeling more predictable after the outcome is known. Looking up the logo and then feeling that the gap was obvious is a different judgment. Confirmation Bias is a preference for evidence that fits a belief already held. The choice task did not score that preference. The second boundary is the cause. Experiment 2 matched the seven icons to seven others on familiarity from Experiment 1: means 3.78 and 3.77, t(12) = 0.11, p = 0.911. During the second experiment, familiarity was 3.65 against 3.82, t(12) = −0.98, p = 0.347, and confidence was 3.24 against 3.48, t(12) = −0.93, p = 0.369. Accuracy was 52.86 percent (standard deviation 15.02) against 82.38 percent (standard deviation 6.07), t(12) = −4.82, p = 4.19 × 10⁻⁴. The degrees of freedom are the 14 images, not the 60 people. MouseView, a cursor aperture meant to resemble looking, found no inspection difference between the two sets, p = 0.143, and none between correct and incorrect choices, p = 0.952. Spectra, the share of foreground pixels, and the color channels did not separate the versions. The paper prints those p values as all above 0.660, all above 0.783, and all above 0.158. Of 58 people who said what they thought the study was about, five, 8.62 percent, mentioned the Mandela Effect or false memory. Of the people who answered the familiarity question about the name, 67.92 percent had heard of it. Those two denominators are not printed as the same number. Do not turn 67.92 percent into a cause. Experiment 3 scraped the top 100 Google images for each of the seven icons. For C-3PO, 51 percent do not show the legs, and 24 percent show the golden-leg version. For Waldo, 28 percent show the wrong feature and 44 percent show the canonical feature. For Volkswagen, those shares are 18 percent and 74 percent. The authors say that for five of the seven, most scraped images are the canonical version, and that Pikachu, Fruit of the Loom, the Monopoly Man, and Curious George are rarely if ever seen in the wrong version. When the cornucopia or the monkey’s tail does appear, they say it is often on a page about the effect, so the memory report comes first. They also say there is no single account. Schema may fill a gap for some icons. Prior exposure may matter for others. A cornucopia, they write, is intuitively atypical for that fruit logo, and Pikachu is a poor fit for a story in which the schema is copied from the character, because the character is the schema. Reported times seen did not differ between the seven and the icons people got right, Wilcoxon z = 0.64, p = 0.523. That item is a bin the person picked, not a viewing log. The third boundary is the drawing, and the export. Experiment 4 asked for 50 drawings of each image. The finished pool is 459 people, mean age 38.1, standard deviation 11.1, not 50 people drawing every icon. Average responses per image were 50.1: 13.2 from people who rated familiarity below 3, and 36.9 from people at 3 or above. The low-familiarity group saw the official picture for three seconds, waited one second, and drew for at least two minutes. The high-familiarity group drew from memory and did not see a picture. Three scorers marked an error only when all three saw it. The rate uses drawings that include the area where the feature could appear. Long-term, the pooled rate is 47.72 percent. The seven image rates average 51.27 percent, standard deviation 36.22. They are not the same summary. By image, long-term: C-3PO 94.44 percent, Waldo 92.86, Curious George 57.69, Volkswagen 43.33, Monopoly Man 31.71, Pikachu 22.73, Fruit of the Loom 16.13. Short-term, the pooled rate is 20.97 percent. The paper prints 100 percent of the C-3PO short-term drawings, 9.09 percent for Curious George, 50 percent for the Monopoly Man, and 33.33 percent for Waldo. The other three icons showed none. It does not print the head count inside the 100 percent. Compared with the single most common drawing error on seven matched icons, the pre-chosen feature was more frequent, mean 39.39 percent against 7.04 percent, F(1, 24) = 11.17, p = 0.003. That control error was chosen because it was the most common, not because it matched a feature named in advance. The samples are adults in the United States, and the icons are the ones those adults would know. The paper says a cultural difference is expected and was not tested. No cell here is a lyric, a household, a brand workshop, or a history class. Priming Effect is a faster word decision. It does not supply these choice shares.

​

Common Misconceptions

Three retellings change the test. One turns a memory error into a changed record. One calls every bad copy by this name. One treats seven icons, and every famous example, as one clean result.

​

Related Concepts

The shared wrong picture sits next to a truth rating, an easy example, a preference for fitting evidence, a foreseeable past, and liking after repetition.

​

One-Line Takeaway