A new peer-reviewed study from Georgetown University and the University of Washington reveals that AI-generated summaries omit critical details and can manipulate human memory, dramatically reducing people’s ability to recall events accurately after reading misleading accounts.
AI-generated summaries have quietly embedded themselves into daily digital life, appearing in search overviews, workplace meeting transcriptions, medical records, and police body camera footage. People rely on them to save time, trusting that machine-condensed text accurately captures the underlying reality of videos and articles. But new research demonstrates that this convenience comes with a steep cognitive cost: misleading AI text can fundamentally rewrite what human witnesses remember about an event.
What Automated Summaries Miss in Animated Accident Tests
To quantify the baseline accuracy of automated text generation, researchers tested two popular models—OpenAI’s ChatGPT and Google’s Gemini—by asking them to summarize short animated videos depicting a car and pedestrian accident. The findings showed that omission is the most pervasive flaw in machine-generated summaries, outpacing explicit fabrication.
Across all tested prompts and models, the summaries omitted an average of 51.6 per cent of central details from the clips. Even more alarming, 95 per cent of the summaries failed to mention the primary event itself: a car colliding with a pedestrian.
Across all prompts and models, the summaries omitted 51.
Study report, AI-Enabled Human Memory Manipulation
The research team pointed out that while hallucination is widely recognized as a primary failure mode for generative AI, systemic omission poses an equally dangerous threat in practical applications.
How Misleading AI Text Distorts Witness Memory
Beyond evaluating machine errors, the Georgetown University and University of Washington team conducted an experiment with 331 human participants to test memory distortion. The participants watched one of two animated videos showing a red car approaching an intersection—featuring either a stop sign or a yield sign—before turning right and striking a pedestrian. Twenty-four to 48 hours later, the participants read a summary of the incident that was either entirely accurate or contained a single altered detail regarding the traffic sign.
When tested on what they had originally witnessed, the difference in recall was stark. Among participants who read an accurate summary, 83.6 per cent answered correctly about the traffic sign. For those who read a misleading summary, correct recall plummeted to 44.8 per cent.
Researchers found that 83.6 per cent of participants who read a summary with accurate information about the video answered correctly, while only 44.
Research team, Georgetown University and University of Washington
Crucially, the experiment revealed that skepticism offers no immunity. Participants were randomly told either that the summary was generated by AI or that it came from a professional human transcriber. Knowing the text came from a machine did not protect readers from adopting false memories, nor did their general trust in or familiarity with artificial intelligence.
Implications for Policing and High-Stakes Legal Settings
The findings carry urgent implications for high-stakes environments where automated summaries are increasingly deployed to speed up administrative workflows.
Mattea Sim, an assistant research professor at the Massive Data Institute within Georgetown’s McCourt School of Public Policy, warned that automated text generation represents an entirely new vector for spreading misinformation.
Eiger expressed particular concern regarding the adoption of video summarization tools in the criminal justice sector without rigorous prior testing. Police departments may be using these technologies unaware of how deeply flawed automated overviews can be. The research findings are scheduled for formal presentation in October 2026 at the Ninth AAAI/ACM Conference on AI, Ethics and Society in Sweden.