US study shows risk of memory distortion from AI summaries — The National

A peer-reviewed study in the United States, released by Georgetown University and the University of Washington, found that inaccurate artificial intelligence-generated summaries of articles and videos can affect how people remember events. As The National reports, such accounts often omit critically important details or inaccurately describe the content.

Omitted details and false memories

The report, AI-Enabled Human Memory Manipulation, states that, on average, summaries omitted 51.6% of key details across all prompts and models used. The researchers tested several AI models, including OpenAI's ChatGPT and Google's Gemini.

The models proved particularly inaccurate when describing an animated video about an accident involving a car and a pedestrian. The researchers found that the content of AI summaries affected participants' subsequent answers about what they had seen.

More current news is available on the UA.News Telegram channel Telegram.

Experiment results

Among participants who read an accurate account of the video, 83.6% answered correctly. Among those who reviewed a misleading summary, 44.8% of respondents gave the correct answer.

University of Washington doctoral student Yael Eiger, who helped lead the study, expressed concern about the potential use of video summarization technologies by police departments without thorough testing and awareness of the risk of errors. Georgetown University assistant research professor Mattea Sim said the findings should help identify situations in which AI summaries should be avoided, especially in critical cases.

The study will be presented in Sweden in October at a conference on artificial intelligence, ethics and society organized by the Association for the Advancement of Artificial Intelligence.