AI summaries of a car accident left witnesses less able to recall what they actually saw

Pattie Maes, a professor at the MIT Media Lab, and her colleagues began with scenes that seemed unlikely to cause trouble: a wedding, a city street and a race. They showed 24 ordinary photographs to 200 people, then brought participants back after a break.

Some saw the original images again. Others saw AI-altered versions, with changed facial expressions, inserted military personnel or added signs of climate change. The experiment was designed to test whether manipulated material could lead people to misremember what had been in the original photographs.

A separate experiment moved from still images to a moving event, and from visual alteration to the increasingly familiar AI summary. Mattea Sim of Georgetown University, working with Yael Eiger of the University of Washington and Tadayoshi Kohno of Georgetown University, studied how misleading information in AI-generated summaries affects memory.

Participants watched a video of a car-pedestrian accident in a laboratory, then read an AI-generated account of what they had watched. Some received accurate information and others misleading information. They then took a memory-recognition test about the original event.

The researchers first prompted large language models to summarise videos and analysed the results. They found frequent mistakes, with the majority of summaries leaving out the video’s central event.

Participants who read a misleading summary were significantly less likely to recall the original accident accurately than participants who read an accurate summary. In effect, a false account supplied after the event altered recall of something the participants had actually witnessed.

That result cuts against the assumption that people supervising AI will reliably catch its errors. The researchers found that memory could be distorted by AI mistakes even where ‘humans-in-the-loop’ are expected to correct them.

The issue has practical force because AI-generated summaries are increasingly used in high-stakes settings, including policing.

The studies also arrive amid a broader argument over where memory resides. A 4E approach in philosophy and cognitive science treats memory as embodied, embedded, extended and enactive, rather than as a process confined within an individual. Under that view, artefacts and technologies can shape memory systems and may partly constitute them.

The Review of Philosophy and Psychology has devoted a special issue to the links among embodiment, memory and digital technology, including generative AI such as large language models and virtual reality. Richard Heersmink and Kourken Michaelian published “Memory and Technology: 4E Perspectives” there on 12 October 2026, observing that no earlier special issue had been dedicated specifically to 4E perspectives on memory and technology.

Another study examined a 68-minute documentary, Tütün Kokan Kadınlar (Women Who Smell of Tobacco), which includes 4 minutes and 13 seconds of AI-generated sequences. It describes the process as “algorithmic memory”, arguing that AI reconstructs the past rather than simply recording it.

In that film, synthetic images create visual continuity for oral testimony where archival material is missing. The study says the sequences organise women’s labour through recurring codes of body, dress and affect, while making that labour more visible and also reducing individual experience to recognisable cultural types.

The paper “AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory” was submitted on 23 Sep 2026.