What are false memories?

Try it

Try the word lists before you read on#

The text further down names the words it tests, so run it first if you’re going to.

What counts as a false memory?#

Roediger and McDermott’s working definition is “either remembering events that never happened, or remembering them quite differently from the way they happened”. The word that causes trouble is “remembering”, because accepting that something happened isn’t the same as having a memory of it. Later researchers separate the two, and the numbers change a lot when they do.

How are false memories made?#

The idea goes back to Bartlett in 1932, who separated rote reproduction from reconstruction, the active filling-in of what’s missing. Roediger and McDermott concluded from their word lists that “all remembering is constructive in nature”. Schacter’s version is that memory isn’t a literal reproduction of the past. Several routes into error have been studied.

Related ideas come to mind

Each list in the demo was built from associates of one word. Underwood proposed in 1965 that seeing “hot” might make you think “cold”, and that the thought is later taken for something you saw. Deese found the association predicts the error: across his 36 lists, the chance that a list’s words called up the missing word correlated 0.87 with how often people wrongly recalled it. His “butterfly” list, whose words didn’t call it up, produced no false recall at all.

Whether people consciously think of the word during the list was open in 1995, and Roediger and McDermott give both possibilities. Two theories are in play. Associative-activation says studying a word activates related words, including ones never presented, and fuzzy-trace says memory forms a verbatim trace and a gist trace, with the error coming from the gist. A 2021 review found the long-term results consistent with both.

Losing track of where a thought came from

You might recall thinking of a word during the list and take that for having heard it. Roediger and McDermott call that a source-monitoring error, one of two possibilities they give, and they lean on an idea from Jacoby and colleagues: the more easily something comes to mind, and the more vivid it is, the more likely it is to be taken for a memory.

Recalling it earlier

In Roediger and McDermott’s second experiment, lists that had been recalled earlier produced more false recognition later, 81% against 72%, and more “remember” answers, 58% against 38%. The same act helped real words too, 79% against 65%.

Information that arrives afterwards

Loftus, Miller and Burns ran 1,242 people through slides of an accident and gave some misleading information afterwards. Those people were less accurate on later tests, and the misleading information mattered more when it came just before the final test than just after the event. In 1995 Loftus and Pickrell wrote that new information acts “like a Trojan horse”, because “we do not detect its influence”.

Suggestion and imagining

In Loftus and Pickrell’s method, people read a relative’s account of the event and were interviewed twice. A mega-analysis of eight such studies classified 30.4% of cases as false memories, and 46.1% when the suggestion was self-relevant, used imagination and had no photo. A smaller study got complete or partial false memories from 10 of 20 people with a doctored photo of a balloon ride. The designs differ, so the two results can’t settle what a photo adds.

Why would memory work this way?

Schacter argues that some of these errors come from useful processes: the errors also reflect keeping useful information about the general themes or meanings of what you met. He cites patient studies and brain imaging, which we don’t go into here. He wrote in 2012 that direct evidence for this view had come only in the previous few years.

The classic evidence, and how sure researchers are#

Roediger and McDermott wrote in 1995 that the surge of interest came mainly from cases in which memories of previously unrecognized abuse were reported in therapy, the “recovered memories” dispute, and this page doesn’t go into that. The “how sure” column below is our own reading of the sources, and nobody has graded these officially.

Word lists: Deese, Roediger and McDermott

Deese published the method in 1959, and it was barely noticed for 36 years. Roediger and McDermott took six of his best lists to a class of 36 students, who heard 12 words at about one every 1.5 seconds and wrote down what they could. They were told not to guess. They wrote the missing word after 40% of lists, while any other wrong word appeared after only 14%.

On a recognition test later, they called the missing words old 84% of the time, nearly the 86% for real words, and unrelated words only 2%. They were sure 58% of the time, against 75% for real words. With 30 new students and longer lists, false recall reached 55%, and on lists followed by arithmetic they called the missing words old 72% of the time and real words 65%. Those students also sorted each “old” answer with the remember/know test. The authors had expected false alarms to come out as knowing, and many came out as remembering.

The demo above cuts down that second experiment, the one in this chart. Its lists are the three printed in the paper’s text: sleep from Experiment 2, chair from Experiment 1, and needle from Deese, as they quote it. Theirs was 16 lists heard from a tape, each followed by two minutes of recall or arithmetic, and the comparison in the demo is their arithmetic lists. The demo has 3 lists you read, with two sums after each that aren’t scored, in shuffled test order, tested on the words at positions 1, 8 and 10 of each list, as they did.

Limits, from the same paper: these were among Deese’s best lists, and the students were at Rice. One student noticed the lists pointed at a missing word, had no false recalls, and was replaced, as the only person with none. They anticipate the objection that word lists and students say little about life outside the lab, and answer that these conditions favoured accuracy. Gallo’s 2010 review asks which aspects generalize, and answers that many opposing processes are involved and need to be separated first.

A 2024 paper calls it the most widely used task for inducing memory distortions. It has also been tried on an unusual group: people with highly superior autobiographical memory fell for related words about as often as controls.

Leading questions: Loftus and Palmer

We couldn’t get this 1974 paper, so the details are as Simply Psychology reports them. Students watched films of traffic accidents and were asked how fast the cars were going when they “smashed”, “collided”, “bumped”, “hit” or “contacted” each other.

A speed estimate could be a guess nudged by the wording, so a second experiment, with 150 students, came back a week later and asked whether they had seen broken glass. There wasn’t any.

How well has it held up? Rolf Zwaan wrote in 2025 that it “has never been directly replicated at scale using modern methods”, and proposed a multi-lab study of about 2,000 people with Loftus’s support, not yet funded when he wrote. Simply Psychology says the verb effect has been reproduced many times, and adds limits: it’s largely confined to peripheral, plausible details, and Yuille and Cutshall found real witnesses far more resistant.

Misinformation: does the memory itself change?

Loftus, Miller and Burns concluded in 1978 that later information is “integrated into the witness’s memory of the event”. McCloskey and Zaragoza answered in 1985 that the usual test can’t show that, and from six experiments with a different test concluded that misleading information “has no effect on memory for the original event”. Simply Psychology reports their test: after reading “wrench” for a hammer, 66% of misled people still chose the hammer over a new item, against 69% of controls. Roediger and McDermott note in 1995 that the size of the effect depends on how it’s tested. We read only the abstracts of the first two, and not what came after them.

Whole events: lost in the mall

Loftus and Pickrell gave 24 people a booklet with four childhood events said to come from a close relative. Three were real. The fourth, being lost in a mall at about five, found by an elderly woman and reunited with family, was built with the relative, who confirmed it hadn’t happened. People wrote what they remembered and were interviewed twice.

Six people, 25%, claimed to remember it fully or partly. In the same study 19 of 24 picked it out as the false story when told one was, and those who accepted it rated it less clear than their real memories, a mean of 2.8 against 6.3 at the first interview, and wrote fewer words about it. The authors called the study “an existence proof”, and wrote: “We make no claims about the percentage of people who might be able to be misled in this way.”

Murphy and colleagues repeated it in 2023 with 123 people and a preregistered plan. Coders rated 35% as having a false memory, 33 partial and 10 full. Asked directly at the end, 14% said they remembered it and a further 52% said they didn’t remember it but believed it had happened.

Andrews and Brewin reanalysed those transcripts in a 2024 paper. The 43 people judged to have a false memory explicitly recalled, on average, fewer than two of the story’s six core details, and 22 of them described something potentially true, distinguishable from the made-up story. The authors then step the rate down: 35% with all 43, 17% without those 22, 11% without a further 7 who didn’t recall being lost, and 4%, which is 5 people, without a further 9 who said at the end that they didn’t remember it.

Wade, Loftus and colleagues replied in 2025. They argue that reports containing remnants of real memories are what a reconstructive memory should produce, and that the original question was whether this can happen at all, so the finding stands “whether the false memory rate in Murphy et al.’s (2023) data is 4%, or 35%, or somewhere in between”. Brewin and Andrews answered that the commentaries misrepresent their work and that concerns about the method persist. We read that answer’s abstract only, and other commentaries not at all.

Earlier, Brewin and Andrews’s 2017 review found some recollective experience in 47% on average across implantation studies, and full memories likely in only 15%. It says “compelling false memories can sometimes be created”, and their 2025 magazine article says it’s “probably much more difficult” than past work suggested. Murphy’s own paper reports that critics of the 1995 study pointed to its lack of a clear definition of “full” and “partial” memories. A 2017 mega-analysis of eight earlier studies classified 30.4% of cases as false memories.

Bartlett’s repeated-recall results, later replicated

In 1995 Roediger and McDermott wrote that Bartlett’s repeated-recall results “never have been successfully replicated”. Four years later Bergman and Roediger ran a closer copy of his method, with delays up to six months, and reported the distortions increasing with time: “Bartlett’s repeated reproduction results can be replicated.”

What do people get wrong about it?#

That a quarter of people can be given a false memory

Wikipedia’s page on the technique says about 25 percent “not only ‘remembered’ the implanted memory but also filled in the missing details”. Simply Psychology concludes that such an event “can be implanted in a substantial minority of ordinary adults”. The paper counted 6 of 24, fully or partly, with short, low-clarity accounts, and made no claim about a percentage for people in general. Later work found the figure depends on how memories are scored: 35%, 14% or 4% in the same study. Simply Psychology does say further down that measurement is contested.

That one word can permanently change a memory

Simply Psychology’s page on the car-crash study says “Even a single word in a question, it turns out, can permanently alter what a person remembers.” The study it describes tested once, a week later, and 68% of the “smashed” group didn’t report glass. The same page says the researchers rejected the response-bias explanation in favour of a genuine change, while also listing demand characteristics as a criticism. It then reports McCloskey and Zaragoza’s conflicting view lower down, that the original memory stays intact.

That false memories always feel certain

Simply Psychology says “Someone with a false memory still feels sure it is true.” It varies. On word lists, people were sure of a word never presented 58% of the time, against 75% for real words. The mall studies measured clarity, which isn’t the same thing: the five people analysed in the original rated the false event 2.8 against 6.3 for their real memories at the first interview, and those coded with a full or partial memory in the replication averaged 3.2 against 6.1 across both interviews. These are group averages from particular tasks, so none of them offers a way to sort a single memory.

Does AI have false memories, or give them?#

On giving them, we read one preprint. In Chan and colleagues’ study, with Loftus as an author, 200 people watched a video of a crime and then answered questions, including five misleading ones, from a survey or a chatbot. The generative chatbot produced over three times as many immediate false memories as the control, and 1.7 times as many as the survey. The number held steady a week later.

Having them is a different question. An AI memory system can store something wrong and bring it back with no sign that it’s wrong, and it can blur where a claim came from, which looks a little like source confusion. That parallel is our reading. Nothing we read shows an experience of remembering behind it.

Counterparts, the memory system whose site this is, keeps what the AI writes down in its own words and brings back what fits. What it wrote can be wrong. The Field Guide follows mechanisms that AI systems do borrow, and the directory compares the systems.

Questions people ask#

Questions about how any of this carries over to AI? Ask in r/AIMemorySystems.