The Misinformation Effect

What it is

A question is not a neutral instrument. The ordinary assumption behind asking someone what they saw is that the memory sits there finished, and that the question is a way of getting at it, so a badly worded question produces a badly worded answer and nothing worse. The misinformation effect is the claim that this is false in an interesting way. The wording gets into the memory. What comes back later is a changed original, rather than the original wearing a distorted label.

The demonstration our note records is the car-crash study. Participants watched a film of a collision and were asked how fast the cars were going when they smashed into each other, or when they hit each other. The smashed wording produced higher speed estimates on the spot, which on its own would be a finding about the moment of answering. A week later the same participants were more likely to report having seen broken glass, and there was no broken glass in the film. Nobody lied. The question had rewritten the memory before anyone tried to retrieve it.

Here is where this page has to be careful about itself. Our note gives no year, no journal, no sample size and no co-authors, and it reaches us through Adam Alter's Drunk Tank Pink rather than through the papers. On this site that combination means unreviewed, and the vocabulary is deliberate: unreviewed records that the material is definitional and unassessed, not that it is shaky. The vault's own note calls this one of the best-replicated findings in cognitive psychology. That sentence is our note's assessment. It is not a replication record anybody here has read, and the difference between those two things is the reason this site publishes an evidence status at all.

Two further cautions belong in the prose rather than in a tag.

The first is the source book. Drunk Tank Pink carries contested on this site, because a large share of the research it relies on failed to replicate: physical warmth priming, money priming, romantic red, Baker-Miller pink. The vault's reliability audit of 31 August 2026 places the misinformation effect in the book's other category, the older non-priming work that it treats as robust, alongside social facilitation, diffusion of responsibility and the mere exposure effect. That placement is a judgment made when the book was read, and a page cannot inherit reliability from a neighbouring paragraph in a book whose headline findings mostly failed.

The second is the attribution. The note names Elizabeth Loftus with medium confidence and records the year as unknown. That is close to the exact shape the vault's verification pass of 1 September 2026 found wrong repeatedly: a confidently named originator with no year attached. Five attributions failed that way in a single tranche, and in every case the credit had gone to the best-known person in the neighbourhood. Loftus is very probably right here, since the work is hers by any account, but our note records no co-authors, and the vault's most frequent correction is precisely the loss of a first author standing next to a famous one. The rule is to credit the first author, and this page cannot yet say who that was.

The mechanism usually proposed underneath the effect is reconsolidation, the finding that recalling a memory reopens it to modification, so a remembered event is rebuilt rather than replayed. That connection is set out on the memory consolidation page, which is robust, and which says explicitly that how much distortion the mechanism produces in practice depends on conditions and belongs to this page rather than to that one. A proposed mechanism is not a second source of evidence for the effect.

In effect

The verb inside the question

The research case is the one this page is built on and is also the one it cannot cite properly. What makes the design worth keeping is the week's delay. A wording effect measured immediately can be explained without touching memory at all: people asked about smashing may simply guess that the researcher expects a bigger number. The false detail reported a week later, the glass that was never in the film, is much harder to explain that way, because it is a positive report of something that did not happen, produced by a single word in a question asked days earlier. Our note records the design and the result. It does not record the numbers, and none are given here.

The witness, the parent, the interviewer

The everyday version needs no film. "Did you see the broken window?" presupposes a broken window, and the person answering has to work around the presupposition rather than through it. Ask a friend how angry someone was and you have already established that they were angry, so the answer negotiates a degree rather than the fact. Ask a child whether the man had a beard and a beard now exists in the conversation. Our note credits this work with reshaping how the courts treat eyewitness testimony, which is the institutional version of the same problem: a witness who has been questioned before arriving has been edited before arriving. On our note's account, the cost of getting this wrong is not that the witness is misled once. It is that the witness now remembers differently, and will remember that way consistently and confidently afterwards.

Feedback forms, and the survey that manufactures its own answer

The commercial version is everywhere and almost never audited. "How much did you enjoy the new checkout?" is a different instrument from "What did you think of the new checkout?", and only one of them has already told the respondent what the answer is about. Five-star widgets that open on five stars, satisfaction scales with more positive than negative options, and post-purchase emails that thank you for a great experience before asking about it are all doing the same work as the verb in the crash question. The vault's own instruction, from the note on Loftus, is that anyone writing customer-survey copy for this site should read this before touching a word of a feedback form. The wider consequence is the one recorded under declared versus revealed preferences: an instrument that shapes the answer it claims merely to measure is not a window onto anything, and a dashboard built on it inherits the wording rather than the truth.

What it does not say

It does not come with a citation attached, and that is the thing to know before using it. No year, no journal, no sample size, no co-authors. The vault's own rule is that a claim not traced to a paper may not be cited in a published article until it has been, so this page exists as a definition and as a work-queue item rather than as a source. The honest position is that we believe the finding, have not checked it, and are saying so.

It does not say that memory is generally unreliable or that testimony is worthless. The claim is narrower and more specific: a particular kind of input, delivered in the wording of a question, gets incorporated. Nothing here measures how often that happens outside a laboratory, or how large the distortion is under ordinary conditions.

It does not inherit its status from anything nearby. Not from the book it reaches us through, whose headline material mostly failed. Not from the reconsolidation work, which supplies a plausible mechanism and no evidence for this effect. Not from Loftus's public standing, which is not a replication record.

And it does not yet carry a verified attribution. The originator field on this page is held at medium confidence with no year, and the vault's most reliable failure mode is exactly this shape. That should be resolved before the concept appears in a published article.

Who developed it

Elizabeth Loftus, on our note's account, established the effect and reshaped how eyewitness testimony is treated in courts. She has no page on this site yet. The vault holds the attribution at medium confidence and records no year and no co-authors, so the specific papers, and whoever stood first on them, remain to be established.

---