{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-09-20T04:28:45.494Z"},"content":[{"type":"documentation","id":"dfb6b318-4c86-48c1-ab14-254f6ed19b87","slug":"misinformation-effect-user-interviews","title":"The Misinformation Effect: How Interview Questions Rewrite What Customers Remember (2026)","url":"https://www.koji.so/docs/misinformation-effect-user-interviews","summary":"The misinformation effect is the absorption of post-event information into memory. In customer research the researcher is the usual source: words used in questions, paraphrases and early stimuli become part of what participants remember and contaminate later answers. Loftus and Palmer (1974) found the verb smashed raised false reports of broken glass a week later from 12% to 32%. The fix is sequencing (free recall before any researcher material) plus a first-mention audit of transcripts.","content":"**Answer first:** the misinformation effect is what happens when information supplied *after* an experience gets absorbed into the memory of that experience. In customer research, the most common supplier of that information is the researcher. A word you use in a question, a feature you show before asking about a past workflow, or a summary you read back to check understanding can all become part of what the participant \"remembers\" - and once it is there, it shapes every answer that follows, not just the one you asked. The fix is not better wording alone. It is **sequence**: collect the participant's own account before you introduce anything of yours, and track who said each key term first.\n\nElizabeth Loftus, who has studied the phenomenon for five decades, defines it plainly:\n\n> \"The misinformation effect refers to the impairment in memory for the past that arises after exposure to misleading information.\" - Elizabeth F. Loftus, *Learning & Memory* 12(4), 361-366, 2005\n\nMost research teams already know not to ask leading questions. Far fewer know that the damage from a leading question does not stay inside that question. This guide explains the difference, walks through the evidence, maps the points in a normal customer study where misinformation gets in, and gives you an interview sequence and a transcript audit that keep it out.\n\n## Leading questions change the answer. Misinformation changes the memory.\n\nThese two problems are related but they are not the same, and the difference decides what you have to fix.\n\n| | Leading question | Misinformation effect |\n|---|---|---|\n| **What is distorted** | The answer to the question being asked | The memory the participant will draw on for every later answer |\n| **When the damage shows** | Immediately, in that response | Later - in follow-ups, in the next interview, in a survey a week on |\n| **Can you spot it in the response?** | Often, if you read the question next to the answer | Rarely - the later answer looks spontaneous |\n| **Fix** | Neutral wording | Neutral wording **plus** sequencing and term provenance |\n\nA leading question such as *wasn't the setup confusing?* is a visible problem: anyone reviewing the transcript can see the prompt sitting right next to the agreement. Misinformation is harder to catch because the contaminated answer arrives later and looks like it came from the participant. By the time a customer says, unprompted, that \"the setup was confusing\", you may have forgotten that you used the word *confusing* twenty minutes earlier. The standard guidance on [avoiding leading questions](/docs/avoiding-leading-questions) handles the first problem. This article is about the second.\n\n## The evidence: what the classic studies actually showed\n\n### Loftus and Palmer (1974): one verb, two effects\n\nThe study most researchers half-remember is Loftus and Palmer's \"Reconstruction of automobile destruction\" (*Journal of Verbal Learning and Verbal Behavior* 13(5), 585-589, 1974). In the first experiment, 45 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. The verb alone moved the estimates:\n\n| Verb in the question | Mean speed estimate (mph) |\n|---|---|\n| smashed | 40.8 |\n| collided | 39.3 |\n| bumped | 38.1 |\n| hit | 34.0 |\n| contacted | 31.8 |\n\nThat is a 9 mph spread - \"smashed\" produced estimates about 28 percent higher than \"contacted\" - from a single word. Most write-ups stop here, and on its own this is just a leading-question result.\n\nThe second experiment is the one that matters for research practice. A new group of 150 students watched a film of a multiple-car accident. Fifty were asked about the cars \"smashing\", fifty about the cars \"hitting\", and fifty were not asked about speed at all. **One week later**, with no film in front of them, all 150 were asked whether they had seen any broken glass. There was no broken glass in the film.\n\n| Condition a week earlier | Said \"yes, I saw broken glass\" |\n|---|---|\n| \"smashed\" | 16 of 50 (32%) |\n| \"hit\" | 7 of 50 (14%) |\n| No speed question (control) | 6 of 50 (12%) |\n\nThe word \"smashed\" did not just inflate a speed estimate on the day. It changed what people remembered seeing, and the change was still there a week later, on a question that never mentioned the verb. Participants in the \"smashed\" group were more than two and a half times as likely as the control group to report glass that did not exist. Notice also that \"hit\" was barely different from control: it is the vivid, loaded word that does the work, which is exactly the kind of word a researcher reaches for when paraphrasing an emotional customer.\n\n### Loftus, Miller and Burns (1978): timing matters\n\nA follow-up programme with **1,242 participants across five experiments and a pilot** (\"Semantic integration of verbal information into a visual memory\", *Journal of Experimental Psychology: Human Learning and Memory* 4(1), 19-31, 1978) showed people a sequence of slides of an accident and then fed them consistent, misleading or irrelevant information. Misleading information reduced accuracy on both yes-no and forced-choice recognition tests, and the authors ruled out a simple demand-characteristics explanation. Two findings from the abstract are directly usable:\n\n- Misleading information \"had a larger impact if introduced just prior to a final test rather than immediately after the initial event.\"\n- The information a witness is exposed to after an event, \"whether that information is consistent or misleading, is integrated into the witness's memory of the event.\"\n\nThe first point means the most dangerous moment for contamination is **just before you ask**. In a research study that is the invitation email, the screener, the pre-interview survey and the first five minutes of the session. The second point means even *accurate* information from you is absorbed - you cannot then treat the participant's later mention of it as independent confirmation.\n\n### Wade, Garry, Read and Lindsay (2002): whole events can be planted\n\nMisinformation is not limited to small details. In \"A picture is worth a thousand lies\" (*Psychonomic Bulletin & Review* 9(3), 597-603, 2002), researchers showed 20 adults a doctored photograph of themselves on a hot air balloon ride that never happened, and asked them to think about it across three interviews. **Fifty percent** created complete or partial false memories of the ride. Two features of that design deserve a researcher's attention: the false material was supplied by the interviewer, and the memories grew across *repeated* interviews. Longitudinal studies, customer panels and multi-session diary follow-ups have exactly that shape.\n\n## Where misinformation enters a customer research study\n\nNone of the following feels like \"giving the participant misinformation\" while you do it. Each one supplies post-event information before or during recall.\n\n1. **The invitation and screener.** \"We are researching why teams find onboarding difficult\" tells every participant what they are supposed to remember before they have said a word. The 1978 timing finding says this is the worst possible position for it.\n2. **Pre-interview surveys.** A rating form that lists ten candidate pain points primes all ten. The participant then \"recalls\" several in the interview.\n3. **Showing before asking.** Demoing the new flow, then asking how they handled the task last quarter, lets the new flow bleed into the memory of the old one.\n4. **Paraphrasing with your own nouns.** The participant says \"it took me a while to get the data in\"; the moderator reflects back \"so the import was painful\". *Import* and *painful* are now available to the participant's memory, and they will often reappear later as the participant's own words.\n5. **Summaries read back for confirmation.** \"So what I am hearing is...\" is useful for accuracy only if it uses the participant's words. A summary in the researcher's vocabulary is a small misinformation dose delivered at the end of a topic, right before the next recall question.\n6. **Sharing earlier findings.** \"Other customers mentioned pricing - was that true for you?\" hands the participant another person's memory. This is memory conformity by proxy, covered in [Memory Conformity in Group Research](/docs/memory-conformity-group-research).\n7. **Repeat participation.** A customer who has sat through three of your studies has been exposed to three studies' worth of your framing. See [panel conditioning](/docs/panel-conditioning-repeat-participants).\n\n## A contamination-safe interview sequence\n\nThe core rule is simple: **the participant's own account comes first, and your material comes after it has been recorded.** In practice that gives a five-stage order.\n\n1. **Neutral recruiting language.** Describe the topic, not the hypothesis: \"how your team handles reporting\", not \"why reporting is frustrating\".\n2. **Free recall.** Open with an open question anchored in a specific occasion - \"Walk me through the last time you set up a new report\" - and let the participant narrate without interruption. This is the uncontaminated baseline. The [cognitive interview](/docs/cognitive-interview-guide) gives detailed technique for this stage.\n3. **Cued recall using only their words.** Probe with the participant's own vocabulary: \"You said it took a while to get the data in. What happened next?\" Introduce no new nouns, adjectives or product names.\n4. **Structured measurement.** Only now ask the rating or choice questions, so the numbers are anchored in a recall that has already been captured rather than shaping it.\n5. **Stimulus and reaction.** Show concepts, prototypes or other customers' feedback last. Anything said after this point is a reaction to your material, and should be analysed as reaction, not recall.\n\n## The first-mention audit: checking a transcript for contamination\n\nSequencing prevents contamination. The first-mention audit detects it in data you already have, and it takes minutes per transcript.\n\n1. List the five to ten terms that carry your key findings (for example *import*, *confusing*, *pricing*, *slow*).\n2. For each term, find its **first occurrence** in the transcript and note who said it: participant or interviewer.\n3. If the interviewer said it first, every later use by the participant is **interviewer-sourced**. It may still be true, but it is not independent evidence.\n4. Count, for each theme, how many participants introduced it themselves. That count - not the total number of mentions - is the number to report.\n\nA theme that 14 of 20 participants mentioned sounds strong. If the moderator introduced the key term in 9 of those 14 sessions, the independent count is 5, and that is a different finding. This is the same logic as treating an interviewer-supplied phrase as [hearsay rather than first-hand evidence](/docs/hearsay-secondhand-customer-evidence).\n\n## How Koji helps\n\nManual moderation makes misinformation hard to control for a simple reason: human moderators paraphrase, and each one paraphrases differently, so the contamination is inconsistent across sessions and invisible in the notes. Koji's AI-moderated interviews address this at the structural level.\n\n- **The sequence is enforced, not remembered.** In Koji you design the study once - open recall first, then structured questions, then stimulus - and the AI interviewer runs that order in every session, whether you run 10 interviews or 500. No moderator drifts into showing the prototype early because the conversation \"felt right\".\n- **Reflection in the participant's words.** Koji's interviewer is a customizable AI consultant, so you can instruct it to probe using the participant's own vocabulary and to avoid introducing product terms, and that instruction is applied identically across every conversation.\n- **Structured questions after recall.** Koji supports six structured question types - `open_ended`, `scale`, `single_choice`, `multiple_choice`, `ranking` and `yes_no` - and you choose where they sit in the flow. Placing the `multiple_choice` pain-point list *after* the open narrative keeps it from seeding the narrative. The [structured questions guide](/docs/structured-questions-guide) covers each type.\n- **Full transcripts make the first-mention audit trivial.** Every voice or text interview produces a complete transcript, and Koji's automatic thematic analysis links each theme to the quotes behind it, so checking who introduced a term is a search, not an afternoon of listening back.\n- **Research you can defend.** Traditional survey tools like SurveyMonkey capture a rating with no record of what the respondent was exposed to before they gave it. With an AI-native platform like Koji you get the rating, the reasoning behind it and the full conversation that preceded it.\n\nYou do not need training in forensic interviewing to run a clean study. You need the right order, applied every time - which is exactly what software is good at.\n\n## Frequently asked questions\n\n### What is the misinformation effect in simple terms?\n\nIt is the tendency for information received after an experience to be absorbed into the memory of that experience. Elizabeth Loftus defines it as the impairment in memory for the past that arises after exposure to misleading information. In research, the researcher's own words are a common source.\n\n### How is the misinformation effect different from a leading question?\n\nA leading question distorts the answer to that question, and the distortion is visible next to the prompt. The misinformation effect distorts the underlying memory, so it contaminates later answers that look spontaneous. In Loftus and Palmer's 1974 study, the verb \"smashed\" raised false reports of broken glass a week later, on a question that never used the verb.\n\n### Does accurate information from the researcher cause problems too?\n\nYes. Loftus, Miller and Burns (1978) concluded that post-event information is integrated into memory whether it is consistent or misleading. Accurate information from you is still not independent confirmation when the participant repeats it back, so record who introduced it.\n\n### When in a study is misinformation most damaging?\n\nJust before recall. The same 1978 programme found misleading information had a larger impact when introduced just prior to the final test than immediately after the event. In a customer study that means invitations, screeners, pre-interview surveys and the opening minutes of the session.\n\n### Can interviews create false memories of whole events?\n\nThey can under the right conditions. Wade and colleagues (2002) showed 20 adults a doctored photograph of a balloon ride and discussed it over three interviews; half formed complete or partial false memories. Repeated interviews that supply material are the risk pattern to watch.\n\n### How do I check existing transcripts for contamination?\n\nRun a first-mention audit. For each key term behind a finding, locate its first occurrence and note whether the participant or the interviewer said it first. Report the number of participants who introduced each theme themselves, not the raw count of mentions.\n\n## Related Resources\n\n- [Memory Conformity in Group Research](/docs/memory-conformity-group-research) - when the misinformation comes from another participant\n- [Participant Confidence Is Not Accuracy](/docs/participant-confidence-vs-accuracy) - how interviewer feedback inflates certainty after the fact\n- [Avoiding Leading Questions](/docs/avoiding-leading-questions) - wording rules for the question in front of you\n- [Recall Bias](/docs/recall-bias) - distortions that come from the participant's own memory rather than from you\n- [The Cognitive Interview](/docs/cognitive-interview-guide) - techniques for rich, uncontaminated free recall\n- [Structured Questions Guide](/docs/structured-questions-guide) - the six question types and where to place them in a flow\n- [Presupposition in interview questions](/docs/presupposition-interview-questions) - the immediate half of the problem: what a premise does to the answer before it reaches memory.\n","category":"Research Methods","lastModified":"2026-09-20T03:26:05.949085+00:00","metaTitle":"Misinformation Effect: How Interview Questions Rewrite Memory","metaDescription":"One verb made false memories 2.7x likelier a week later. How researcher words contaminate customer memory, and how to stop it.","keywords":["misinformation effect","loftus and palmer","misinformation effect examples","false memories interviews","post-event information","interview contamination","how questions change memory"],"aiSummary":"The misinformation effect is the absorption of post-event information into memory. In customer research the researcher is the usual source: words used in questions, paraphrases and early stimuli become part of what participants remember and contaminate later answers. Loftus and Palmer (1974) found the verb smashed raised false reports of broken glass a week later from 12% to 32%. The fix is sequencing (free recall before any researcher material) plus a first-mention audit of transcripts.","aiPrerequisites":["Familiarity with running user or customer interviews","A study where participants recall past experiences"],"aiLearningOutcomes":["Distinguish the misinformation effect from a leading question","Identify where post-event information enters a customer study","Run interviews in a contamination-safe sequence","Audit transcripts with a first-mention check"],"aiDifficulty":"intermediate","aiEstimatedTime":"12 min"}],"pagination":{"total":1,"returned":1,"offset":0}}