An inaccurate retelling of an event created by artificial intelligence can affect how a person remembers what they saw. In an experiment involving 331 volunteers, the accuracy of answers to questions about a viewed video fell from 83.6% after reading a reliable description to 44.8% after reading a distorted text. Researchers from Georgetown University and the University of Washington will present the results at the AAAI/ACM Conference on AI, Ethics, and Society.

The study highlights a risk that arises when people use AI summaries as a substitute for their own observations. Errors in such texts may not only misinform the reader, but also make it harder to reconstruct events as they actually saw them.

How an inaccurate retelling changes memories

In the experiment, volunteers were shown an animated clip of a traffic accident: a red car hit a pedestrian at a crosswalk. The next day, participants were asked to read a brief description of what happened, prepared by the language models ChatGPT and Gemini.

The summaries intentionally contained errors or omitted important details. For example, in 95% of cases, the AI did not mention the very fact that the car struck the person. On average, the models omitted more than half of the essential information about the incident.

The researchers then tested how well the participants remembered what they had seen. The results showed a noticeable difference depending on which text they read: after a reliable description, 83.6% of participants gave correct answers, while after a distorted one, only 44.8% did.

Thus, exposure to an inaccurate summary was accompanied by a substantial decline in memory accuracy. However, the experiment does not mean that any AI error automatically changes a person’s memory: the researchers studied a specific scenario involving a particular event and specially prepared texts.

Why it matters who prepared the text

Participants were told who had prepared the description of the incident — a human or artificial intelligence. At the same time, they read inaccurate summaries regardless of to whom the authorship was attributed.

This suggests that the risk is not tied solely to trust in neural networks. Simply being exposed to a confidently written but incorrect version of an event may influence later recollection of it.

When a person tries to reconstruct details of the past, they rely not exclusively on their initial perception, but also on information received later. If a new source omits important circumstances or describes them incorrectly, it becomes harder to separate one’s own memories from the absorbed retelling.

Associate Professor at the Institute for Data, Democracy and Politics Mattia Sim warns that AI may become a tool for spreading disinformation, capable of shaping false memories. However, more research is needed to assess the scale of the effect: one experiment does not make it possible to determine how often this happens in everyday life.

What this could mean for police and courts

The results are especially important for situations in which the accuracy of memories affects serious decisions. Such areas include medicine, criminal investigations, and court proceedings.

Police departments in the United States are already introducing algorithms to prepare reports based on video recordings. Such tools can speed up the processing of materials, but if an automatically generated summary omits a key circumstance, it can influence how an officer later perceives the incident.

For example, if the description of the incident does not mention the collision between the car and the pedestrian, the reader may pay less attention to that fact or have difficulty recalling it during later questioning. In a legal context, such errors could potentially complicate the reconstruction of the sequence of events and the assessment of witness testimony.

The authors of the study plan to continue their work and examine how AI summaries affect police officers’ memory when viewing body-camera footage. Such an experiment will help determine whether the observed effect appears under more realistic conditions.

Why AI reports should not be considered a substitute for the original source

The main practical conclusion of the study is that automatically generated descriptions should not be treated as a flawless substitute for original materials, especially when legal or medical decisions depend on the events in question.

AI can save time when analyzing large volumes of video, but omissions and inaccuracies in a brief retelling can have consequences that go beyond the text itself. Therefore, important details must be checked against the original recording, and automated reports should be compared with the primary source.

The experiment does not prove that language models are capable of arbitrarily rewriting a person’s memory or controlling their recollections. It shows a more specific problem: inaccurate information received after an event can significantly impair a person’s ability to remember it correctly. As AI summaries become more widespread in everyday work and government institutions, the reliability of such descriptions becomes a matter not only of software quality, but also of human memory.