Misleading AI Summaries Distort Eyewitness Memory

AI-generated summaries can change what people remember about an event they watched. In a controlled study from Georgetown University and the University of Washington, participants who read an inaccurate account of an animated traffic incident later recalled key details correctly less often than those who read an accurate account: 44.8% versus 83.6%.

That difference matters because a summary error can outlast the summary itself. Someone may later rely on a memory that has already been shaped by an incorrect description, rather than simply consult a flawed document.

The experiment separated summary errors from memory effects

The researchers first tested OpenAI’s ChatGPT and Google’s Gemini on animated traffic videos. Across the tested summaries, the models omitted an average of 51.6% of central events. In 95% of tested iterations, they left out the collision between a car and a pedestrian.

For the memory experiment, 331 participants watched an animated car-pedestrian incident, then read a narrative summary 24 to 48 hours later. The summaries contained either accurate or altered details. Researchers subsequently tested participants’ memory of the original event.

The findings do not show that every video summary will produce this degree of error. They do show a concrete failure mode in this task: a summary that misstates or omits details can affect a reader’s later recall, not just their understanding of the text in front of them.

An AI label did not act as a safeguard

Participants’ acceptance of false memories was similar whether they believed the summary had been written by an AI system or a human transcriber. The report also found that participants’ prior familiarity with and trust in AI did not protect against memory contamination. Disclosure alone, in other words, did not prevent the observed effect.

That finding complicates the idea that a person reviewing an AI summary is enough to catch its mistakes. In this experiment, people were not just evaluating a document; they had first seen the event and then encountered an inaccurate account of it. The summary could become part of the material they used to reconstruct what happened.

What the results mean for high-stakes summaries

The researchers raise policing as a context where this risk deserves attention, including the potential use of summaries of body-worn camera footage. But the study used animated scenarios, not real police footage, and did not test whether officers’ reports or courtroom testimony change. The researchers described real-world body-camera footage as future work.

For teams using summaries of consequential events, the practical distinction is between a digest and a record. The study suggests an inaccurate digest may influence a human witness’s later recollection, so keeping the source footage available does not by itself undo that influence. Whether the same pattern appears with real footage and professional workflows remains to be established.

FAQs

In the experiment, participants who read inaccurate summaries were less likely to recall key details of the animated event correctly than those who read accurate summaries, suggesting errors can shape later recollection. The study tested animated scenarios, not real-world eyewitness testimony.