To investigate how misleading information in AI-generated summaries affects recall of events.
Approach:
Summary Analysis: AI models ChatGPT-5.5 and Gemini 2.5 Flash-Lite summarized 2 animated videos of a car-pedestrian accident, producing 20 summaries.
Online Experiment: 360 participants watched one of the videos and then read an AI-generated summary that was either accurate or misleading regarding a traffic sign.
Recall Assessment: Participants' ability to identify the traffic sign was measured, along with the influence of the summary's source (AI vs. human) on recall.
Key Findings:
Correct recall of traffic signs was 84% for consistent summaries and 45% for misleading summaries, according to the study.
The misinformation effect was consistent across both videos, regardless of the traffic sign seen, as reported by the authors.
The source label (AI or human) did not significantly affect recall outcomes, according to the findings.
Participants' trust in AI and frequency of AI use did not modify the misinformation effect, as noted in the study.
Errors were present in all AI-generated summaries, with significant omissions and inaccuracies identified by the researchers.
Interpretation:
Errors in AI-generated summaries can distort human memory.
Limitations:
The study involved only 2 short animated videos and 2 AI models, limiting generalizability.
The source-label comparison reflected beliefs about authorship rather than independent AI and human summaries.
Conclusion:
Errors in AI-generated summaries can distort human memory.
Increases in apnea-hypopnea events were greatest among patients with an elevated baseline index, while evidence supporting mouth taping remained preliminary.
