A study of simulated witness interviews found more false memories after misleading chatbot questions. The wording of a question can introduce details.
Suppose you are trying to remember a disagreement at work. You tell a chatbot that a colleague left the room, and it asks what they said before storming out. The word “storming” adds something you did not report. If the conversation continues from there, your account may become more definite than your memory actually is.
This example is invented, but the underlying concern has been tested. In an experiment using a crime video, researchers examined what happened when people later received misleading questions from an AI chatbot. The chatbot condition produced more false memories than the control condition. The study concerned a staged event and a particular interview setup, not every conversation people have with AI.
The risk is serious enough to justify a clear boundary: a chatbot should not guess the missing details of an event you are trying to document. It can arrange notes you already have, but a plausible addition is still an addition. A readable account is not necessarily an accurate one.
A 2024 author preprint by Chan and colleagues studied 200 people who watched a crime video. People then had different follow-up tasks. One group got misleading questions from a chatbot. The chatbot condition produced more false memories than the control condition. The study also included a later check. It tested leading interviews about a staged event, not all everyday chats.
The version reviewed here is a preprint, which means it had not completed peer review in that version. Read the authors' paper. The important concern is not that a chatbot can erase your past. It is that details supplied after an event can become mixed with what you think you saw.
Compare “What did the person carry?” with “What did the person do with the bag?” The second question assumes a bag was present. Someone uncertain about the event may start trying to remember its use instead of questioning that assumption. A follow-up can therefore change the task from describing an event to evaluating a detail supplied by the interviewer. That is why preserving the difference between what the person reported and what the question introduced is important.
That can happen in ordinary notes. A tool may turn He left quickly into He ran from the room. It may turn They argued into She shouted at him. The change makes the story more vivid, but it adds details. Once an added detail appears in several drafts, it may become difficult to remember whether it came from the original account or from later editing. Saving the first version gives you a record to compare against rather than relying on that distinction staying clear in memory.
If you need to record an event, start with your own plain notes. What did you see? What did you hear? What are you unsure about? Keep those parts separate. For example: I heard a loud sound. I did not see what made it. Someone later said a glass fell. That is clearer than I heard the glass fall if you never saw a glass. Save the first version. Do not overwrite it with a smoother account. A rough note with a clear source can be more useful than a polished story that mixes sight, guess, and hearsay.
This is an editorial suggestion for keeping records clear. It is not a legal interview method or a guarantee that memory will stay exact.
Editing often rewards completeness, but missing information should remain missing when you are recording an event. A fluent transition that supplies an unreported action may improve readability at the cost of accuracy. Write I do not recall the color. Write I am not sure which person spoke first. A useful account can contain both clear facts and honest gaps. If the tool adds a detail, ask where it came from. Was it in your note? Was it in a separate record? Or was it a guess made to complete the story? Remove guesses that have been turned into facts.
Do not keep asking until the tool gives a version that feels right. Familiarity is a poor test when the same exchange is creating that familiarity.
A photo, a time-stamped message, or a receipt may help you check part of an event. Note what each record shows. It may settle the time without settling what someone meant. Suppose a message proves you arrived at three. It does not prove you were calm, angry, or alone. Let each piece of evidence answer the question it can answer. If a record conflicts with your first note, keep the difference visible. I first recalled four; the receipt says three-fifteen. That is a better record than silently making the first note agree.
The goal is not to defend your memory at all costs. It is to keep track of why the account changed.
Keep the original account and compare any edited version against it. Pay particular attention to changes in actions, timing, certainty, and descriptions of other people's intentions. “Left the room” and “stormed out” can lead a reader to different conclusions even though the rest of the sentence is unchanged. The study does not prove that a chatbot can rewrite someone's entire memory. Its finding is narrower and useful: misleading follow-up questions can contaminate an account of a recent event. When accuracy matters, preserving an uncertainty is better than accepting a detail just because it makes the story easier to follow.
