Distinct paths to false memory revealed in hundreds of narrative recalls

Memory distortions emerge from a complex interplay between prior knowledge and ongoing experience—dynamics which are not readily provoked in controlled laboratory experiments. Here we investigate naturally occurring memory distortions using the largest known dataset of narrative recall, comprising hundreds of spoken recollections. Using large language models (LLMs), we developed an automated pipeline to detect and classify spontaneous false memories. Across two validation experiments, we demonstrate that human-AI agreement matches inter-human reliability in detecting and cataloging memory distortions. We show that false memories reflect two distinct phenomena which are driven by separable semantic factors: similarity to prototypical narrative patterns drives factual errors (distortions of actual content), whereas contextual surprise drives confabulations (entirely fabricated details). Through this combination of large-scale naturalistic data and AI-powered automation tools, we reveal memory processes that controlled laboratory paradigms cannot easily capture and illuminate the complex dynamics of human (mis)remembering in real-world contexts.