Researchers at Georgetown University and the University of Washington examined how summaries produced by artificial intelligence affect human memory. The study was motivated by the widespread use of AI to condense videos and events, and by the recognition that models can make mistakes — producing hallucinations, misinterpretations, or omitting important details.
Method: analyzing AI summaries and an animated-video experiment
The team first analyzed summaries generated by ChatGPT and Gemini and found frequent errors: models omitted critical facts, misinterpreted information, and sometimes hallucinated inaccurate details. The examined summaries omitted about half of the important details on average.
In the experimental part, participants watched short animated videos showing a red car approaching an intersection marked either with a Stop sign or a Yield sign; which sign each participant saw was assigned at random. In the videos the car turned right and struck a pedestrian.
After 24–48 hours, participants were randomly given an AI-generated summary to read. Some summaries were accurate, others contained at least one inaccuracy. Participants were told whether the summary they received was written by an AI or a human, but in fact every summary text was generated entirely by ChatGPT.
Results: a single false claim greatly reduces correct recall
Participants were asked whether the car in the video had approached a Stop sign or a Yield sign. Among those who read an accurate summary, 83.6% answered correctly. By contrast, only 44.8% of participants who read a misleading summary reported the correct detail. A single incorrect piece of information reduced correct recall by more than 38 percentage points. The researchers noted this effect was not influenced by whether participants believed the summary was produced by a human or by AI.
Implications and next steps
The authors conclude that AI can generate misinformation even without malicious intent, and that such errors can substantially alter human memory. They warn this is particularly concerning for high-stakes contexts — for example, police video summaries — where AI summarization might be used without rigorous testing.
For future work, the researchers plan to study real-world footage such as police body-camera recordings and the AI-generated police reports derived from them, to assess how these summaries might reshape memories of police–civilian interactions and perceptions of incidents. "I think we should be thinking very deeply and critically about how we should use AI to summarize information, especially in high-stakes situations," said Mattea Sim, the lead researcher on the study.
The paper is currently available as an arXiv preprint, and the authors call for further research to map risks in real-world applications.
