Understanding spontaneous false memory in the naturalistic recall of narratives

Poster Session C - Sunday, March 30, 2025, 5:00 – 7:00 pm EDT, Back Bay Ballroom/Republic Ballroom

Phoebe Chen¹ (hc2896@nyu.edu), Omri Raccah², Todd M. Gureckis¹, David Poeppel^1,4, Vy A. Vo³; ¹New York University, ²Yale University, ³Intel Labs, Intel Corporation, Hillsboro, OR, ⁴Center for Language, Music, and Emotion, NYU & Max Planck Institute, Frankfurt, Germany

Human memory has traditionally relied on highly controlled, trial-based paradigms. A growing body of work is now investigating memory function under naturalistic conditions, primarily focusing on the recall of accurate details in audiovisual and spoken narratives– as such, memory errors for narrative content remain largely unexplored. This study examines the factors that drive spontaneous false memories in the natural retelling of stories. Leveraging a recently published dataset comprising hundreds of verbal recalls of four spoken stories (Raccah*, Chen*, et al., 2024), we employed in-context learning with large language models (LLMs) to detect memory errors. First, we automatically segmented the story into distinct events and prompted GPT-4o to align recall sentences with corresponding story events. Next, we used GPT-4o to identify and categorize memory errors in each recall sentence. The model learned to perform the scoring by relying on examples of memory errors (factual conflicts, confabulations) and reasonable inferences based on the story. We validated the LLM scoring by using human ratings for a subsample of participant recollections (average human-AI agreement: 0.73; average inter-rater agreement: 0.72). Next, we correlated the rate of false recollections for each event in the story with established factors known to influence memory performance. Our results revealed that contextual surprise – i.e., the model-estimated likelihood of an event given all previous events – significantly predicts the false memory rate (β=0.18, p=0.009). This work demonstrates the potential of LLMs in evaluating errors in naturalistic recollections and uncovers contextual factors that determine spontaneous false memories.

Topic Area: LONG-TERM MEMORY: Episodic