A summary containing just one false piece of information can alter the memory we retain of a scene we previously witnessed.
Researchers from Georgetown University and the University of Washington tested this effect using an animated scene depicting an accident. A car approaches an intersection with either a stop sign or a yield sign. It then turns and hits a pedestrian. Participants had to remember what they had actually seen.
Between 24 and 48 hours later, they read a summary of the scene generated with ChatGPT. Depending on the group, the text correctly indicated the sign or mentioned the other one. The researchers then asked participants to answer based on their memory of the video, rather than the summary.
The result is clear. Among the 328 people included in the analysis, 83.6% of those who had read the correct information in the ChatGPT text identified the correct sign. With misleading information, this proportion fell to 44.8%. The text read after the video had therefore altered the answer concerning a specific detail of the scene.
One might think that a warning about the text’s origin protects against this effect. Participants were specifically told that the summary came either from an AI or from a human transcriber. In reality, they all read text produced with ChatGPT. The “AI” label did not significantly reduce the influence of the erroneous information.
The experiment draws on a phenomenon known for several decades: information received after an event can alter the memory of that event. The researchers therefore do not show that AI creates a new weakness in the brain. They find that an automated summary can become a new channel through which incorrect information affects memory.
The team also examined twenty summaries produced from two short videos by ChatGPT and Gemini. All contained errors according to their analysis method. On average, 51.6% of the details considered central were omitted. Nineteen out of twenty summaries did not even mention the collision between the car and the pedestrian.
The researchers now want to study situations from the real world, particularly summaries of videos recorded by police officers. This step will make it possible to test whether the same effect appears with richer scenes, longer documents, and decisions with concrete consequences.