Yes, That's My Handwriting

On the morning after the space shuttle Challenger broke apart, a psychologist at Emory University handed his students a questionnaire. Where were you when you heard? Who told you? What were you doing? They wrote it down. He put the answers away.

Two and a half years later he found forty-four of those students and asked again. Then he scored each new account against the one that same person had written the morning after, on a scale of nought to seven. The average came out at 2.95 (Neisser & Harsch, 1992). About a quarter of the accounts, Neisser wrote afterwards, were entirely wrong, and plenty of those were delivered with high confidence.

Entirely wrong does not mean blank: every detail they scored had changed. They had ordered, specific memories of a different morning.

Then Ulric Neisser did the thing that made the study famous. He gave them back their own questionnaires, in their own handwriting, dated January 1986.

They read them and kept the new version. One student looked at the page and said: "Yes, that's my handwriting — but I still remember it this other way!"

That is the difficulty. Nobody here has a blank or a fog. They have a detailed account, held sincerely, sitting next to written proof that it is wrong, and it wins.

Something had happened in those two and a half years, and nobody had lied. One of the changes shows the shape of it. The morning after the explosion, nine of forty-two students said they had first heard the news on television. Two years on, nineteen of the same forty-two said television (Greenberg, 2004). They were remembering real footage. It just was not the footage of that morning. They had all watched it over and over in the following weeks, and the memory of watching slid backwards into the slot where the original moment used to sit.

This happens to people with unlimited access to the record. George W. Bush described hearing about the September 11th attacks on three public occasions, and twice he recalled watching the first plane hit the tower on live television. Nobody watched that. No footage of the first impact was broadcast that morning (Greenberg, 2004). The same error, running in the same direction, in a man with every resource in the world for checking.

The game the English called Russian Scandal

The idea that remembering is a construction job rather than a playback is about a century old, and it starts with a Cambridge psychologist reading a story aloud.

Frederic Bartlett gave people "The War of the Ghosts", a Native American folk tale chosen because it made no sense to an English undergraduate, and asked them to reproduce it. In one version of the method each reproduction is handed on to the next person in a chain, so the story travels through six or seven heads. Bartlett gave that method its name, serial reproduction, though the technique had been in use among British anthropologists for decades already, a provenance he later tidied up in the telling (Ost et al., 2022). He took it up, by his own account, after the mathematician Norbert Wiener asked whether he could "do something with 'Russian Scandal', as we used to call it". Russian Scandal was the English parlour game. Russian-speaking children play the same game and call it испорченный телефон, the spoiled telephone.

The results are the ones everyone half-knows. The story got shorter, tidier and steadily more English, and unfamiliar things turned into familiar ones.

Modern chains show which part goes first. In a 2022 study, eighty first-year students in sixteen chains of five lost specific details far faster than they lost the gist. The wording collapsed at the first retelling and then largely levelled off, which led those authors to doubt that the chain is what drives the distortion at all. That asymmetry, the shape holding while the particulars evaporate, turns out to explain a good deal of the disagreement in this literature about how bad memory is.

Here is the part I did not expect, and it cuts against the metaphor I am building. Bartlett's subjects did not tell the story to anybody. They read it and wrote their versions out, alone. He knew this mattered and said so in print: "To write out a story which has been read is a very different matter from retailing to auditors a story which has been heard. The social stimulus, which is the main determinant of the form in the latter case, is almost absent from the former." The founding study of reconstructive memory is not, strictly speaking, a study of retelling at all.

For actual mouths, you want the rumour researchers. Gordon Allport and Leo Postman ran chains where one person described a picture aloud and the description travelled down a line of six or seven listeners. About 70 per cent of the details vanished in the course of five or six transmissions, they reported, "even when virtually no time lapse intervenes" (Allport & Postman, 1945). Be precise about what that number is: details lost, under instructions to be accurate, in front of an audience, with no motive to embroider. Allport and Postman listed five ways their laboratory failed to reproduce a real rumour, and noted that four of the five should have made their subjects more accurate.

Their most famous result concerns a picture of a subway car containing a white man holding a razor and a Black man. Somewhere along many of those chains, the razor moved into the Black man's hand.

That finding has been retold in textbooks, review articles and courtrooms for eighty years, and the retellings are wrong. Two researchers went back to the source in 1987 and found the drift (Treadway & McCloskey, 1987). The published claim had become that half of the observers, after "a brief look" at the picture, remembered the razor in the Black man's hand. In the actual procedure nobody in the chain ever saw the picture. One member of the audience described it aloud while looking at it, and everyone downstream worked from that description. Allport and Postman's own claim was that in more than half of their experiments the razor ended up in the wrong hand at some point in the chain. They ran no control condition, so whether the migration shows racial stereotyping or something duller was never actually tested. Treadway and McCloskey traced the corrupted version into sworn expert testimony. Then, correcting it, they misstated a procedural detail themselves.

A study about how retelling deforms a story, deformed by retelling, in the literature that studies retelling, three levels deep. I have not found a better demonstration of anything.

Retelling is not recalling

The reason a retold memory drifts is that telling a story and recovering one are different jobs, and the first has an audience.

In a 1978 experiment, people read a description of somebody and then wrote a summary for a listener whose opinion of that person they had been told in advance. The summaries leaned toward the listener, which is unsurprising. What matters is what happened next: when asked later to reproduce the original description, they reproduced their own slant. Thirty-four per cent of the reproductions were distorted in the positive direction when the audience liked the target, against four per cent when the audience disliked him (Higgins & Rholes, 1978). The control condition is where it gets strange. People who were told the audience's attitude, read the same material, and expected to communicate but never actually wrote anything showed no such bias. Knowing what your listener thinks does nothing to your memory. The damage comes from having said it out loud to them.

People are fairly aware of this. Asked to log their own conversations for a month, thirty-three students labelled 61 per cent of their retellings as distorted in some way, by exaggeration, omission, minimising or addition, while calling only 42 per cent inaccurate (Marsh & Tversky, 2004). That is people owning up to spin, which is a different animal from a measured error rate, and the authors think it understates things, since people prefer to see themselves as truthful. The general principle, as Elizabeth Marsh put it, is that "conversational retellings depend upon the speaker's goals, the audience, and the social context more generally" in a way laboratory recall does not (Marsh, 2007).

Other people's versions get in as well. Pairs of witnesses who watched slightly different videos of the same crime and then discussed it went on to report details they had never seen: 71 per cent of them mentioned at least one item that existed only in the other person's version (Gabbert, Memon & Allan, 2003). That number gets quoted a lot, usually as though seven in ten people had their memories overwritten. The follow-up work complicates it. When a later study asked people where each detail had come from, most of them could say correctly. They had reported the borrowed detail knowing it was borrowed, because the situation invited it. Reading the other person's written report was about as contaminating as talking to them (Bodner et al., 2009). Conversation shapes testimony more reliably than it rewrites memory, and the two get conflated constantly.

What conversation does do is decide which parts survive. When a group talks through a shared event, the parts that get mentioned are strengthened for everyone in the room, and related parts that go unmentioned fade, in the listener as well as the speaker. This has been shown for word lists first and then for people's memories of September 11th (Coman, Manier & Hirst, 2009). A family, a regiment or a diaspora talking regularly about the same past is converging on one version and letting the rest go.

None of that quite says what "a rumour you keep retelling yourself" implies, and I would rather deal with the gap than talk past it. Very little of this work is about a person revisiting their own past in private. Bartlett's subjects wrote alone. Allport and Postman passed a description between strangers. Gabbert's witnesses changed what they said more readily than what they held. The narrower claim that survives is that producing an account changes the account you will produce next time, and what shapes it is whoever you are producing it for. There is almost always somebody.

Where it stops being a parlour game

Of the first 375 DNA exonerations recorded in the United States between 1989 and 2020, 69 per cent involved a mistaken eyewitness identification (Innocence Project). That dataset is closed; the organisation stopped counting nationally at 375, so the figure describes a closed archive and the count has not moved since. The broader picture is less lopsided. The National Registry of Exonerations, which records every known exoneration obtained by any means rather than DNA cases alone, listed 3,842 exonerations as of 25 July 2026, of which 1,039, about 27 per cent, involved mistaken witness identification. Mistaken identification is not even the leading factor in that larger set. Perjury or false accusation is, in roughly two thirds of cases.

Both numbers are true and they describe different populations. DNA exonerations are heavily stranger sexual assaults, the crime where an identification is often the entire case and where biological evidence exists to overturn it later. That is the selection, and it is why the eyewitness share is so high there.

Confessions do it too. Twenty-nine per cent of those 375 cases involved a false confession by the defendant or by a co-defendant who implicated them. In the laboratory version, students accused of crashing a computer by pressing a key they had not pressed signed a false confession 69 per cent of the time. Signing is compliance and proves little on its own. But 28 per cent came to believe they had done it, and 9 per cent went on to confabulate details of doing it. In the harshest condition, where the pace was fast and a false witness claimed to have seen it, every single participant signed and about two thirds internalised it. In the gentlest condition, a third signed and nobody believed it (Kassin & Kiechel, replicated by Horselenberg et al., 2003).

The result that bothers me most is a small one. After a witness picks someone out of a lineup, the officer says something like "Good, you identified the suspect." That one sentence changes what the witness remembers about their own state of mind at the moment of picking. Among witnesses who had chosen the wrong person, 50 per cent of those given confirming feedback later rated their certainty at 6 or 7 on a seven-point scale, against 15 per cent of those told they had picked the wrong man (Wells & Bradfield, 1998, which had no control condition, only the two feedback groups). Across nineteen tests and some seven thousand participants, confirming feedback inflated retrospective certainty by nearly a full standard deviation (Steblay, Wells & Douglass, 2014).

The question they were answering was how sure they had felt at the time. Two of the thirteen things measured were completely immune to the feedback, which I find reassuring in a strange way: how long the culprit had been in view, and how far away he had been. Feedback rewrites judgements and leaves durations and distances alone.

In a related demonstration, nobody in the room believed any of this had happened. All of the lineup administrators and 95 per cent of the witnesses denied that any feedback had been given.

Having a better memory does not protect you

There are people who can tell you what they had for lunch on a Tuesday in 1987, and be right. Highly superior autobiographical memory is rare and real, and it is the obvious place to look for immunity.

Twenty of them were tested against thirty-eight matched controls on five standard distortion tasks (Patihis et al., 2013). On a word-list task designed to produce false recognitions, they falsely recognised 70.3 per cent of words they had never seen. The controls managed 70.8. Asked whether they had seen footage of United Flight 93 crashing in Pennsylvania, which does not exist, a fifth of them said yes, statistically indistinguishable from the controls.

The flip side belongs in the same paragraph, because the popular version always cuts it. On the same task these people were better at recognising the words that had actually been shown, 76.6 per cent against 64.8. Their storage really is extraordinary. It buys no protection at the reconstruction end. With twenty participants these are underpowered null results, so all I can claim is that the likeliest candidate group for immunity distorted at ordinary rates. Ruling immunity out would take a bigger sample.

Being clever helps a little. In a study of over four hundred students, the best cognitive predictor of resisting misinformation accounted for roughly a tenth of the variation between people (Zhu et al., 2010). Individual differences are enormous, from people who swallowed almost none of the misinformation to people who swallowed almost all of it. Intelligence is in there. It is not a shield.

The errors can also be shared across strangers, down to the same wrong detail. A study of forty pop-culture images found seven, or six once a statistical error in the analysis was corrected, that large numbers of Americans misremember identically: a monocle on the Monopoly man, a black tip on Pikachu's tail, two gold legs on C-3PO. Across ten thousand random splits of the sample, which wrong version people picked correlated at 0.88 between halves (Prasad & Bainbridge, 2022). The recognition task has been replicated independently (Transparent Replications). Note the denominator, though, since nobody ever quotes it: for thirty-two of the forty images, people were significantly more likely to be right than wrong. Shared false memory is a striking exception rather than the normal condition.

The part where I had it backwards

I came into this assuming the moral was that confidence is worthless, since that is the version everybody repeats. It is not what the evidence says, and the correction is the most interesting thing I found.

In 2017 John Wixted and Gary Wells reviewed the field and argued close to the opposite (Wixted & Wells, 2017). Under what they call pristine conditions, and specifically at the first identification, high confidence is a strong signal of accuracy. In a Houston police field study they estimated that sixty-nine of the seventy-two high-confidence identifications of a suspect were correct, with a model putting high-confidence accuracy near 97 per cent and low-confidence accuracy closer to a coin flip.

Those conditions are strict. A fair lineup, only one suspect in it, no hint about which face is the suspect, an administrator who does not know either, and confidence recorded verbatim right then, before anyone says a word about it. Most jurisdictions still do not guarantee all five. The percentages are model estimates rather than observed outcomes, and they rest on an assumption about how often the police have the right man in the first place.

The shape of the argument survives all that, and it reorganises the courtroom material above. Their reading of the exonerations is that the disaster started with a hesitant identification, which then got turned into a confident one somewhere between the police station and the trial, by feedback and repetition and the months of retelling in between. Where the trial record preserves it, the certainty on the stand was often built on a first identification that had been shaky.

This is a live argument rather than a settled result. Elizabeth Loftus and colleagues have pushed back on exactly that point: for most exonerees, nobody wrote down the initial confidence, so the archive cannot prove what it is being asked to prove (Berkowitz et al., 2022). A 2025 meta-analysis of seven field studies covering nearly two thousand real witnesses found that among high-confidence identifications made to an administrator who did not know the answer, about one in eight landed on a filler, a person known to be innocent (Fitzgerald et al., 2025). The same Houston data yields very different error rates depending on an assumption neither camp can settle.

There is a second correction, and for daily life it matters more. Memory is not generally bad. When seventy-four adults freely described two real, staged experiences at delays running from days to nearly three and a half years, between 93 and 95 per cent of the checkable details they volunteered were accurate (Diamond et al., 2020). The amount they could produce fell sharply over the years. The proportion that was right held up. Three quarters of them still got at least one thing wrong, which is the other half of the sentence, and neither half should be dropped.

That also answers an objection this essay's own opening invites. Neisser's students look catastrophic and these people look fine because they were asked different questions. A questionnaire with a line for who told you and a line for what you were doing forces an answer onto every line, and a line whose contents have gone will get filled with something plausible. Ask people instead to hand over whatever they still have, and what they hand over is mostly right. The gist survives, the particulars go, and a form that demands particulars will collect inventions.

So the accurate summary is narrower and less quotable than the one I started with. As three researchers who have spent careers arguing about this put it, memory is "clearly malleable but not unreliable, under normal circumstances and in the absence of contamination or prolonged suggestion by psychologists, therapists, or anybody else" (Brewin, Andrews & Mickes, 2020). The contamination is the story. A memory left alone is usually decent. Handle it repeatedly, in company, under pressure, with somebody telling you that you got it right, and you are dealing with a different object.

Which inverts the intuition I started with. The memories you would defend hardest are, by definition, the ones you have gone over most often, in company, with something riding on how they land. Neglect protects. The story you never tell is in better condition than the one you tell every year.

What none of this licenses

There is a way of using this research that I want to shut off before going any further, because it has done damage in actual courtrooms.

"Memory is unreliable" is a slogan, and it has been used to tell people that a witness who tells a story slightly differently the second time must be making it up. The British Psychological Society's guidance to the courts is careful in both directions. Memories "are not a record of the events themselves" and should not be compared to a video recording. And, in the same document: "some inconsistencies in memory across time (both for trauma and non-traumatic information) are common and this does not necessarily imply that the memory or the event is fabricated" (BPS, 2008).

On abuse specifically, the field's position has been stable for thirty years and is more careful than either side's partisans. Most people abused as children remember what happened to them. Memories recovered after a long gap can be true, false, or a mixture, and the risk rises sharply when they surface through suggestive techniques. Constructing convincing memories of things that did not happen is possible (APA). The researchers most associated with false-memory work, Elizabeth Loftus among them, put their own line in print: be careful "not to discredit genuine cases of sexual trauma" and "to take corroborated claims of such trauma seriously" (Otgaar et al., reply to Brewin, 2021).

The base rates point the same way. In a prospective study, 129 women with hospital-documented childhood sexual abuse were interviewed seventeen years later, and 38 per cent did not report the abuse that was in their own medical record (Williams, 1995). What goes wrong here is usually silence. The measured rate of false allegations of sexual assault sits somewhere around 2 to 10 per cent (Lisak et al., 2010).

Memory being reconstructive is a reason to record things early and handle witnesses carefully. It is not a reason to disbelieve people.

A country doing it out loud

Everything above happens inside one head, or in a conversation between two. Something that rhymes with it happens at the scale of a nation. What follows is an analogy rather than evidence, because polling measures what people are willing to say and not what they hold. It earns its place anyway, because at this scale the rehearsal is public, it is dated, and somebody can be seen putting a thumb on it.

Ask Russians whether they regret the collapse of the Soviet Union and the answer moves like weather. Levada recorded 66 per cent in March 1992, 75 per cent in December 2000, then 49 per cent in December 2012, the only reading under half, then back up to 66 per cent in November 2018 and 63 per cent in November 2021 (Levada). Between the low and the following peak, seventeen points in six years, nothing new was learned about 1991.

What people say they miss also drifts, though less tidily than the story wants. Among those who regret it, the share naming broken ties with relatives and friends fell from a high of 38 per cent in 2007 to 20 per cent in 2021. The economic reason, the destruction of a single economic system, has topped the list in seven of the ten waves, though not in 2006, 2012 or 2014, when the loss of great-power status did; it read 60 per cent in 1999 and has sat between 48 and 55 ever since, ending at 49. The great-power reason itself stood at 46 per cent in 2021, ten points up on 2018 and still below its readings in 2006, 2012 and 2014. Quote only the pair that fits and you get a clean march from the personal to the national. The table does not show one.

Set the numbers against what the same organisation was recording at the time. In autumn 1989, 73 per cent of people in Soviet Russia said they felt a lack of basic foodstuffs often or constantly, and 84 per cent said the supply of food and goods had noticeably worsened over the previous two or three years. Almost half of those asked in 1990 said famine in the country was possible the following year (Levada). By 2020, the words most associated with the Soviet era were stability and confidence in tomorrow (Levada).

Reputations move the same way. Asked to name the most outstanding people of all times and nations, 12 per cent of respondents named Stalin in December 1989, and 35 per cent by 1999 (Levada). In Levada's ranking of outstanding figures he has come first since 2012 (Levada). Meanwhile the state pollster VCIOM, in a survey fielded in February 2026 and released the following month, found that 14 per cent of Russians born in 2001 or later regret the loss of the USSR (VCIOM). They are missing a country that ceased to exist before they were born.

Some qualifications, because this could very easily turn into a cheap shot in two directions.

These are surveys of stated attitudes, not measurements of anybody's memory. A person can revise their verdict on the 1970s without a single autobiographical memory changing. The numbers moved; saying they were moved is a causal claim the polling does not support. Nor is this a Russian condition. Ask who contributed most to defeating Germany and 95 per cent of Russians say the USSR, while 59 per cent of Americans say the United States and 13 per cent say the USSR (Levada, YouGov). Same war. The Americans have been telling themselves one story about it for eighty years and the Russians another, and the two do not meet anywhere.

And the nostalgia is not stupid, which is the thing I would least like misread. Russian output per head fell by more than 40 per cent from 1989 and did not recover its 1989 level until 2007 (World Bank). A third of the population had incomes below the subsistence minimum in 1992 (official series). People who remember the 1990s as ruin are remembering correctly. What the contemporaneous record contradicts is the memory of the late Soviet period as a time of stability and confidence in tomorrow.

Where the collective case differs from the individual one is that somebody can lean on it. Since 2020 the Russian constitution has said that "diminishing the significance of the people's feat in the defence of the Fatherland is not permitted" (Article 67.1). In December 2021 the Supreme Court ordered the liquidation of Memorial, the organisation that had spent three decades assembling the archive of Soviet repression (BBC), and in April 2026 the movement was declared extremist outright (Human Rights Watch). The polling I have been quoting comes from an organisation the Justice Ministry placed on the register of foreign agents in 2016, a designation it disputes, and which it must announce on every page it publishes (Meduza). Legislating the conclusion and closing the archive is the national version of the officer saying "good, you identified the suspect."

Estonia's case works differently, and that is why it is worth having. Nothing about the Levada pattern applies: no drift, no revision over time. One object stood still while two communities rehearsed incompatible accounts of what it was, until the accounts collided hard enough to move it. On the night of 26 April 2007, work at the site of a Soviet war memorial in Tallinn brought a crowd, and rioting followed; in the small hours an emergency cabinet session ordered the Bronze Soldier removed at once. One man was killed, dozens were injured and more than three hundred people were detained (RFE/RL). For Estonia's ethnic Russians, about a quarter of the population, the statue and the thirteen Red Army graves beneath it stood for the sacrifice the Soviet Union made in helping to liberate Europe. For many Estonians, whose country was annexed by the USSR in 1940 and got its independence back only in 1991, the same bronze meant occupation (Postimees). The bronze itself never changed. Two communities had spent fifty years telling themselves what it stood for, and neither version was invented.

The system is not broken

The obvious ending here is that we are all hopelessly unreliable. I think that ending is wrong.

Daniel Schacter sorted memory's failures into seven categories and then said the thing worth remembering about them: they "are by-products of otherwise adaptive features of memory" rather than faults in the design (Schacter, 1999, updated 2022). A memory that could not be recombined would be no use for the job it also does, which is imagining what has not happened yet (Schacter & Addis, 2007).

Retrieval makes the same point more sharply. Students who tested themselves on a passage recalled 61 per cent of it a week later. Students who instead read it repeatedly, going over it more than four times as often, recalled 40 per cent, and the re-readers were the more confident of the two groups (Roediger & Karpicke, 2006). That is the same split the lineup feedback produced, certainty rising while accuracy does not, arrived at here with nobody applying any pressure at all. Retrieval reshapes a memory. What you pull up gets stronger, and so does anything you got wrong on the way up; what you never pull up fades.

A few practical things follow. Write down what happened while it is fresh, and keep the note, because the note will be better than you will. Be more careful with the story you have told the most times, not less. If a group of you is trying to work out what happened, take everyone's account separately before you compare, because once you have compared, you have merged. And when somebody remembers an evening differently from you, both of you being certain settles nothing at all, since certainty is what the process produces on its way to the answer.

The students in that room in 1988 were not fools and they were not lying. They were doing the ordinary thing, which is telling a story they had told before. The only unusual part of their situation was that somebody had kept the first draft.