Georgetown, University of Washington: AI summaries distort memory

The study sits atop a long line of psychology research on the "misinformation effect," in which false information given after an event alters memory of that event. The team's twist was to ask whether that effect applies even when the source is an AI system people already know can be wrong.

The experiment separated two things. The AI summaries themselves were poor: across 20 summaries, every one contained errors, and the collision at the heart of the video was missing from 19 of them. Separately, when 328 U.S. adults watched the video and later read a summary with the road signs swapped, their recall of the signs dropped sharply compared with those who read an accurate summary. Recall of details the summary never mentioned stayed high.

What the results suggest is that the danger may not be limited to AI asking leading questions; a summary read after the fact could be enough. Because telling readers the text came from AI made no significant difference, the outcome hinges on whether such effects show up outside the lab, which the team intends to test with real police body-camera footage.