Study finds AI summaries significantly distort human memory recall.

A recent study demonstrates that artificial intelligence summaries can fundamentally alter what individuals remember about specific events. The research highlights a significant cognitive cost associated with the convenience of automated text generation, showing that these tools do not merely condense information but actively reshape human recall.

The Misinformation Effect in Action

Participants viewed a 25-second animated clip depicting a red car approaching an intersection where a pedestrian stepped into the road. The vehicle turned and struck the pedestrian, who then fell, rose, and confronted the driver. Two nearly identical versions were used: one featured a stop sign at the intersection, while the other displayed a yield sign. This single visual difference served as the critical variable for testing memory reconstruction.

This experimental design draws on psychological research spanning more than five decades, which establishes that memory functions not as a static video file but as a reconstructive process. Information encountered after an event can integrate into this reconstruction, a phenomenon known as the misinformation effect. The study sought to determine how AI interacts with this established cognitive vulnerability.

Impact on Recall Accuracy

After viewing the video, participants returned 24 to 48 hours later to read a 21-sentence AI-generated summary of the incident. For half the group, the summary accurately described the traffic sign they had seen; for the other half, researchers deliberately altered the description. A participant who witnessed a stop sign might read that the car approached a yield sign. Instructions required them to answer based solely on their original visual memory, disregarding the text.

The results revealed a stark divergence in accuracy. Among those reading the correct summary, 83.6% recalled the traffic sign properly. Conversely, only 44.8% of those exposed to the misleading summary retained the correct detail. This drop was not due to general forgetfulness. Participants answered 91.4% of other memory questions correctly, including recalling that the accident occurred during daylight, a detail absent from the summary. More than 96% remembered the time of day accurately, indicating robust overall retention except where AI provided conflicting data.

Transparency Does Not Mitigate Error

Researchers tested whether disclosing the source of the summary would protect participants from misinformation. Some subjects were told the text was generated by AI, while others were informed it came from a professional human transcriber. In reality, all texts were AI-generated. The disclosure had negligible impact on resistance to false details.

Additionally, the study examined whether individual trust levels or frequency of AI chatbot usage influenced susceptibility. No significant evidence emerged linking either factor to protection against misleading summaries. Mattea Sim, a study author, stated in a Georgetown University press release that AI represents a new method of delivering misinformation capable of creating false memories. She emphasized the need for critical evaluation of AI summarization, particularly in high-stakes environments.

Omission Over Hallucination

The team also evaluated the fidelity of the AI systems themselves, using ChatGPT-5.5 and Gemini 2.5 Flash-Lite to describe the two video clips. Each model generated five summaries per video, totaling 20 outputs. Every summary contained errors, ranging from incorrect details to fabricated elements. However, omission proved the most prevalent failure mode, with models excluding 51.6% of central event details on average.

Nineteen of the 20 summaries failed to mention that the car hit the pedestrian at all. While hallucination is often cited as the primary risk of AI, this study identifies omission as a potentially more pervasive issue. Even without inventing facts, AI summaries can mislead by removing critical context. Unlike transcription errors confined to the page, summary inaccuracies can become embedded in the reader’s mental narrative.

The findings suggest that as users increasingly rely on summaries for lengthy or complex material, they face a trade-off between speed and detail preservation. The study, published in arXiv, has not yet undergone peer review.