Elizabeth Loftus and John Palmer showed forty-five students films of traffic accidents, then asked them to estimate how fast the cars had been going. The question was identical for every participant except for one word.
About how fast were the cars going when they hit each other? The average estimate was thirty-four miles an hour. When they smashed into each other? Forty-one. Collided, bumped and contacted produced estimates in between, in the order you would expect.
That is interesting but explicable as a response bias — the verb might simply have influenced how the students chose to phrase an uncertain estimate, without changing anything about what they remembered.
The second experiment closed that off. A new group watched a film of a car accident and was asked the speed question with either smashed or hit, or was not asked about speed at all. Then they were sent away. A week later they returned and were asked a set of questions including this one: did you see any broken glass?
There was no broken glass in the film.
Of the students who had been asked the hit question, fourteen percent reported seeing broken glass. Of those asked the smashed question, thirty-two percent did.
The verb had not merely shifted an estimate at the time. A week later it was producing a visual detail that had never existed, in a memory the witness experienced as their own.
This is the misinformation effect, and Loftus spent the following decades establishing its scope: information encountered after an event is integrated into the memory of the event, and once integrated, it is not marked as having a different source. The witness cannot tell which parts of their recollection came from the world and which from the question. Confidence does not distinguish them either; in several of Loftus's studies false details were reported with as much confidence as true ones, and in some later work with more.
The mechanism sits in the reconstructive nature of episodic memory. Recall is not playback. Each retrieval assembles the memory from fragments plus schema plus context, and during that assembly the trace is labile — the reconsolidation literature suggests a window in which the reassembled version, including anything added, is what gets stored back.
The operational implication for anyone conducting an interview is severe, and it is the whole reason Part 4 of this guide exists in the shape it does. An interviewer who asks leading questions is not merely risking a bad answer in this session. They are permanently altering the witness's memory, and any subsequent interview will retrieve the altered version with increased confidence. Contamination is not recoverable.
It also gives an offensive technique that is used routinely in cross-examination, in hostile journalism and in coercive relationships: describe the event to the person in your preferred terms, repeatedly, and wait. The chapter on gaslighting is this mechanism applied domestically over months.
The defense is to record your own account of anything consequential immediately, in your own words, before anyone else has described it to you. That document is the only version that cannot be edited afterwards, including by you.
Cross-ref: Part 4 cognitive interview