Home — Computing and AI

In briefA joint study by researchers at Georgetown University and the University of Washington has found that AI-generated summaries containing incorrect information can alter people's memories and shape their perception of what actually happened, even when they are told the content was produced by AI.

1. AI-generated summaries are appearing with growing frequency — in news previews, workplace meeting transcriptions and high-stakes contexts such as police body camera footage — even as research continues to show that AI systems can produce misleading or inaccurate outputs. The new study, titled "AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory," has been published and is scheduled to be presented at the Ninth AAAI/ACM Conference on AI, Ethics and Society (AIES), a peer-reviewed academic conference taking place in October 2026.

2. The study was led by Mattea Sim, an assistant research professor at the Massive Data Institute within Georgetown's McCourt School of Public Policy. Her co-authors were Yoshi Kohno, the McDevitt Chair in Computer Science, Ethics, and Society and a professor in Georgetown's Department of Computer Science and Center for Digital Ethics, and Yael Eiger, a Ph.D. candidate at the University of Washington. Kohno described the work as part of a broader effort at Georgetown to examine the relationship between AI and humans through the combined lens of psychology and computer science.

3. Two-Part Experimental Design

4. The research team took two distinct approaches. In the first, they analysed AI summary outputs from OpenAI's ChatGPT and Google's Gemini to measure how often errors occurred and what form those errors took. The study found that AI summaries can contain frequent mistakes, omit important details, misrepresent information and hallucinate inaccurate content. Eiger, whose research centres on technology in the carceral system, said she was struck by the poor quality of the summaries and expressed concern that police departments may be deploying video summarisation tools without adequate testing or awareness of how errors in those summaries could cause harm.

5. For the experimental portion, participants were asked to watch one of two animated videos — created as part of a recent replication study that updated materials from a classic human memory study — depicting a red car approaching an intersection marked with either a stop sign or a yield sign, with viewers randomly assigned to see one or the other. In each video the car turns right, collides with a pedestrian who then stands up, and the driver exits the vehicle to meet the pedestrian. The researchers noted that no actual police body camera footage was used.

6. Between 24 and 48 hours later, participants were randomly assigned to read an AI-generated summary that was either accurate or contained a deliberately altered detail. Before reading, they were also told that the summary had been written by either an AI or a human transcriber. In practice, all summaries were generated by ChatGPT, with portions of the model's output stitched together in some cases to maintain consistent experimental conditions. A total of 331 people completed both parts of the study.

7. Memory Altered Even With AI Disclosure

8. Participants who read a misleading summary were significantly less likely to accurately recall the original event compared with those who read an accurate one. Crucially, the researchers found that informing participants that the summary was AI-generated did not prevent memory distortion. Sim and her colleagues write in the study that while humans in the loop are often expected to catch AI errors, their findings suggest that human memory can instead be warped by those errors, and that AI has the potential to generate misinformation even without any adversarial intent.

9. "AI is a new method of delivering misinformation, and it has the potential to create false memories for people who are reading that information," Sim said, adding that careful thought is needed about whether and how AI should be used to summarise information, particularly in high-stakes settings. Kohno said he hopes the findings prompt greater awareness that AI outputs can directly influence how people perceive events, especially in consequential fields such as policing or medicine.

10. Calls for Transparency and Further Research

11. Eiger said she would like to see more audits of AI summarisation technology. She noted that while the study used the most current consumer-facing versions of large language models, the authors did not have access to enterprise versions of AI built for organisational use. She also pointed to what she described as limited public transparency around the exact models being deployed as AI adoption accelerates across sectors. For future research, the team intends to move beyond animated videos and examine real-world cases, including how AI-generated police reports based on body camera footage might shape memories of police-civilian encounters and perceptions of what took place.

Based on reporting by Mirage News

Also in this issue

• QuEra Survey: 45% of Firms Now Prioritize Quantum Fault-Tolerance RoadmapsQuantum Zeitgeist
• DOE Report Sets Three-Phase Roadmap Toward Error-Corrected Quantum ComputingQuantum Zeitgeist
• AI Infrastructure Spending Forecast to Hit $769 Billion in 2026 Amid Safety WorriesET Enterprise AI
• DOE Publishes National Quantum Computing Roadmap After Broad Stakeholder ProcessFermilab (.gov)
• Microsoft Opens Maryland Quantum Center to Host DARPA System TestingHPCwire