A research team from Georgetown University and the University of Washington has reported that reading incorrect summaries created by generative AI can cause readers to recall events in accordance with misinformation, even when they have directly watched the original video.
SecurityGeorgetown UniversityUniversity of Washington
AI Misleading Summaries May Distort Human Memory, Experiment Reveals
This article is a translation. Read the Japanese original
The research team conducted experiments using OpenAI's ChatGPT-5.5 and Google's Gemini 2.5 Flash-Lite to summarize videos of traffic accidents. Despite providing instructions to the AI such as "be factual and do not speculate," the generated summaries contained numerous errors. Specifically, it was confirmed that the central event of the video, "a car struck a pedestrian," was omitted in 19 out of 20 summaries.
Next, to investigate the impact of incorrect summaries on human memory, an experiment was conducted with 328 adults. After showing participants videos of traffic accidents, they were asked to read summaries between 24 and 48 hours later. In the group that read accurate summaries, 83.6% answered correctly regarding the content of signs, whereas the accuracy rate for the group that read incorrect summaries dropped to 44.8%.
On the other hand, for peripheral memories not mentioned in the summaries, such as the "time of the accident," many participants maintained correct answers, revealing that memory of the video as a whole was not lost. The research team pointed out that simply recognizing that AI makes mistakes may not be sufficient to prevent the impact of misinformation on memory, regardless of whether it is explained that the summary was created by AI.
Sources
- AIの間違った要約を読むと「自分で見た出来事」まで間違って記憶することが実験で判明 （GIGAZINE、2026-09-29）