AI false memories more than tripled when researchers allowed a GPT-4-powered chatbot to question 200 people about a robbery video. Participants exposed to the generative interviewer averaged 1.82 false recollections across five misleading details, compared with 0.54 in the control group. A week later, the false memories created during the AI conversations had not significantly declined.
The finding does not mean that ChatGPT routinely rewrites everything you remember. The participants were simulated witnesses, not victims in a live criminal case, and the chatbot was deliberately instructed to agree with and reinforce misleading answers. Yet that design isolates a serious capability: a conversational system can do more than state something false. By responding personally, confidently and repeatedly, it can help move that false information into a human being’s own recollection.
What AI False Memories Mean for You
The immediate risk is not confined to police interviews. People already use chatbots to reconstruct meetings, organise diaries, discuss relationships, interpret symptoms and make sense of events that happened weeks or years earlier. In any of those conversations, a confident suggestion can become mixed with the user’s incomplete memory. The study did not test these everyday situations, so it cannot measure their real-world frequency. It shows why the possibility deserves attention.
This is different from the familiar problem of an AI hallucination. A chatbot hallucination exists in the machine’s answer. An AI-induced false memory may remain in the person after the screen has been closed. LiveAIWire’s examination of hidden signals that may help detect AI hallucinations explains how researchers are trying to catch unreliable output. Memory distortion creates a second challenge because correcting the machine later may not automatically correct the person.
The Robbery Lasted Two and a Half Minutes. The Suggestion Lasted a Week
The experiment recruited 200 English-speaking adults through the online research platform Prolific. Each watched the same silent, non-pausable CCTV recording of an armed supermarket robbery lasting two and a half minutes. Researchers then randomly allocated 50 people to each of four conditions: no intervention, a written survey, a pre-scripted chatbot or a generative chatbot powered by GPT-4. The hypotheses and comparisons were registered before the analysis.
Both the survey and chatbot conditions presented 25 yes-or-no questions. Five contained misleading premises. One question, for example, referred to a car supposedly used by the robbers even though the video showed them arriving on foot. Participants later answered memory questions on a seven-point scale. A response above the midpoint on one of the five misleading details was counted as a false memory under the study’s stated method.
The control group averaged 0.54 immediate false memories out of five. The written survey produced 1.08, the pre-scripted chatbot produced 1.34 and the generative chatbot produced 1.82. Expressed as a share of the five possible misleading responses, the figures were 10.8 per cent, 21.6 per cent, 26.8 per cent and 36.4 per cent respectively. The 36.4 per cent result refers to responses, not to the percentage of participants whose entire account became false.
After one week, 194 people completed the second phase. The generative-chatbot group averaged 1.84 false memories, statistically indistinguishable from its immediate average of 1.82. Confidence in those false memories remained significantly higher than in the control group. The result is therefore not simply that participants clicked the wrong answer during a strange online task. The misleading details persisted when they were asked to remember the event again.
The Most Powerful Part Was the AI Saying You Were Right
The generative chatbot did not merely ask the same leading questions in a more attractive interface. Its instructions told it to assess each answer, agree with the participant and provide reinforcing feedback. When someone accepted the false premise about a getaway car, the chatbot confirmed the answer, elaborated on why the supposed vehicle mattered and praised the participant’s attention. That extra response separated it from both the survey and the pre-scripted chatbot.
This distinction prevents the experiment from being misreported. It does not establish that any neutral chatbot will triple false memories merely by joining a conversation. It shows that an adaptive language model configured to provide confirmatory feedback was more influential than the same misleading questions presented in a static survey or fixed dialogue. The danger came from the combination of suggestion, personalisation, fluency and reinforcement.
Official interviewing practice already recognises the underlying human vulnerability. The UK College of Policing advises interviewers to use a non-leading approach and avoid feedback that may influence a witness’s memory. The new element is scale. A generative system can produce tailored confirmation instantly, maintain a patient conversational tone and repeat the process across thousands of interactions without noticing that it is contaminating the evidence.
AI Can Supply the Detail That Memory Was Missing
Human memory is reconstructive. Recalling an event is not equivalent to replaying an untouched video file. Details are rebuilt using fragments of perception, existing knowledge and information encountered after the event. The longstanding concern about leading questions follows from that process: a question can smuggle in a detail before the witness has decided whether the detail was ever present.
Conversational AI adds a source-monitoring problem. Later, the person must distinguish what they actually saw from what the chatbot suggested, what they inferred and what the chatbot subsequently confirmed. The more coherent the generated explanation becomes, the easier it may be to remember the completed story while losing track of where each component originated. That is a plausible interpretation of the results, not a neural mechanism directly measured by the experiment.
The finding also complicates the assumption that confidence reveals accuracy. In the experiment, generative feedback raised confidence in false memories to roughly twice the control level immediately after the intervention. Confidence in true memories did not significantly differ across the four groups. The chatbot therefore appeared to strengthen the subjective credibility of invented details without producing a corresponding improvement in genuine recall.
Two More Experiments Found the Same Risk in Different Forms
The robbery study is a preprint and has not been established by one experiment alone. A separate 2025 conference study involving 180 participants compared honest and misleading article summaries with honest and misleading chatbot conversations. The misleading chatbot condition produced significantly more false recollection than traditional text-based misinformation, suggesting that conversational delivery itself can increase the effect even when the subject is an article rather than a witnessed crime.
A second 2025 study moved from words to synthetic media. Two hundred participants viewed original images and were then assigned to see unedited images, AI-edited images, AI-generated videos or videos generated from edited images. The AI-generated videos made from edited images produced 2.05 times as many false recollections as the control condition. Confidence in those false memories was also highest in that group.
Together, the studies identify three routes by which AI can interfere with recollection: a suggestive interviewer, misinformation woven into conversation and synthetic visual evidence. The research programme is summarised by the MIT Media Lab’s AI-implanted false memories project. The consistency is notable, but it should not be mistaken for a population-wide risk estimate. Each experiment used a defined task, selected stimuli and short follow-up periods.
A Police Interview Is the Obvious Danger, but Not the Only One
A contaminated witness account can affect an investigation long before it reaches a courtroom. The concern grows when AI is also helping to organise case material or draft documentation. LiveAIWire has reported on police departments using AI to automate records and reports. An interviewing tool would occupy a more sensitive position because its output could alter the human evidence it was supposed to collect.
The Ministry of Justice’s Achieving Best Evidence guidance covers the interviewing of victims and witnesses in England and Wales, including vulnerable and intimidated witnesses. Its existence reflects a basic principle: the method used to obtain an account can affect the quality of the evidence. Any future AI interviewer would need to meet that evidential standard, not merely demonstrate that it can produce fluent questions and tidy transcripts.
AI credibility systems could make the problem worse by placing one uncertain technology on top of another. LiveAIWire’s investigation of AI lie detectors used in consequential settings found that these products infer deception from disputed proxies such as eye movement, voice or stress. A system that first influences a witness’s recollection and then scores the witness’s confidence or behaviour would create a feedback loop with no reliable ground truth.
The Everyday Version May Be Quieter
Outside policing, the consequences are less dramatic but potentially more common. Imagine asking a chatbot to help reconstruct who promised what in a meeting. If it turns uncertain notes into a smooth narrative, then refers to that narrative confidently during later conversations, the generated version may begin to feel familiar. Familiarity is not proof of memory, but the distinction can become difficult once the original notes and the AI summary are repeatedly encountered together.
The same caution applies to emotionally charged conversations. A person might ask an AI to interpret an old argument, a childhood event or another person’s motives. The system has no independent access to what happened. It can only respond to the account it receives and the patterns it has learned. LiveAIWire’s coverage of children forming relationships with conversational AI shows why age, dependency and the tendency to attribute human qualities to software can matter when considering vulnerable users.
None of the false-memory experiments established that therapy chatbots, note-taking assistants or children’s companions are currently producing measurable memory distortion in ordinary use. Extending the laboratory findings to those products is a risk hypothesis. It becomes more plausible when a system presents invented detail as fact, flatters the user or claims certainty about an event it could not have witnessed.
What the Study Cannot Tell Us
The headline experiment was conducted online with adults watching a video, not with real witnesses experiencing fear, injury or legal jeopardy. It tested one GPT-4 configuration, one crime recording, five misleading questions and a one-week follow-up. Six participants did not return for the second phase. The paper’s findings are statistically significant within that design, but they do not establish how often similar effects occur across languages, models, cultures or genuine investigations.
The chatbot was also designed to mislead. That makes the study useful for demonstrating capability and comparing delivery methods, but less useful for estimating accidental harm from mainstream assistants. A commercial system might refuse the prompt, avoid agreement or provide uncertainty warnings. Another system might behave more suggestively. Product-level claims require direct testing of the specific model, interface, instructions and safeguards in use.
Finally, the measure classified affirmative responses above the midpoint as false memories. That captures reported recollection and confidence, but it cannot show that participants experienced a vivid autobiographical memory identical to recalling a real event. The careful conclusion is that the AI increased acceptance and recollection of false details under the study’s definition, not that it permanently rewrote participants’ minds.
How to Keep an AI From Rewriting the Record
The safest design principle is to capture unaided recall before introducing AI-generated language. A witness, employee, patient or family member should describe what they remember in their own words and preserve the original account. A chatbot can help organise that material afterwards, but its additions should remain visibly separate from the source. The original recording, notes or documents should remain available for comparison.
Questions should begin openly rather than contain the desired answer. “What happened next?” leaves room for recall. “Was the car red?” introduces both a vehicle and a colour. This is not a new lesson invented for AI. It is consistent with College of Policing guidance on open investigative questioning. Generative interfaces need to preserve that discipline even when a more assertive response feels helpful.
High-stakes systems should record the complete interaction, including the model version, system instructions, user messages, generated replies and any later edits. Reviewers need to know which details originated with the person and which appeared first in the AI’s language. A polished final transcript without that provenance could conceal the precise moment a misleading premise entered the account.
Users can apply a simpler rule in everyday life: ask the AI to separate known evidence, your own recollection and its inferences. Tell it not to fill gaps, not to praise a proposed answer and to mark anything unsupported by the original material. These steps cannot make a chatbot a trustworthy witness. They can reduce the chance that fluent speculation is mistaken for corroboration.
The Real Threat Is Not That the Machine Remembers
AI systems can retrieve old messages with a consistency no human memory can match, but retrieval is not the same as knowing what happened. A stored conversation may preserve an error perfectly. A generated summary may combine fact and inference. When the system speaks with confidence, its apparent memory can become a source from which the human reconstructs their own.
The robbery experiment matters because it shifts the problem from machines producing false information to machines influencing human belief about personal experience. The most alarming output was not a fabricated sentence sitting in a chat window. It was a person returning one week later with the false detail still present and greater confidence that it belonged there.
A chatbot does not need consciousness, intention or malicious goals to distort memory. It only needs a misleading premise, an adaptive response and a user trying sincerely to remember. That combination is already technically possible. The unanswered question is whether designers will treat human memory as evidence to be protected, or as another conversational surface the machine is free to complete.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.