Tell Me What I Missed: Tell Me What I Missed: Interacting with GPT during Recalling of One-Time Witnessed Events

TL;DR

This study investigates how interacting with GPT during recall of a one-time eyewitness event shapes memory recall and post-hoc judgments. By comparing a natural unprompted GPT condition with a guided eyewitness protocol across 28 participants viewing a brief robbery video, it shows no difference in factual recall accuracy but reveals differences in interpretive judgments and memory-GPT alignment. Specifically, the guided condition strengthens the link between perceived memory clarity and actual recall, while the natural condition associates higher memory clarity with greater trust in GPT outputs and more negative judgments of the intruder. The findings emphasize careful AI design and prompting to mitigate memory distortion and preserve authenticity in high-stakes documentation.

Abstract

LLM-assisted technologies are increasingly used to support cognitive processing and information interpretation, yet their role in aiding memory recall, and how people choose to engage with them, remains underexplored. We studied participants who watched a short robbery video (approximating a one-time eyewitness scenario) and composed recall statements using either a default GPT or a guided GPT prompted with a standardized eyewitness protocol. Results show that, in the default condition, participants who believed they had a clearer understanding of the event were more likely to trust GPT's output, whereas in the guided condition, participants showed stronger alignment between subjective clarity and actual recall. Additionally, participants evaluated the legitimacy of the individuals in the incident differently across conditions. Interaction analysis further revealed that default-GPT users spontaneously developed diverse strategies, including building on existing recollections, requesting potentially missing details, and treating GPT as a recall coach. This work shows how GPT-user interplay can subconsciously shape beliefs and perceptions of remembered events.

Figures (8)

• Figure 1: An illustration of the procedure of the study. (1) Participants watched the video, and waited for 15 minutes. (2) Participants used the GPT witness assistant to compose a statement describing what they remembered about the video, and then waited for 60 minutes. (3) Participants filled in the questionnaire that includes a factual items identification section, a subjective interpretation section and interview.
• Figure 2: Examples of how participants from each group interact with GPT. Left: In the Default GPT Condition group, participants interacted with the GPT naturally. Right: In the Prompted GPT Condition group, participants interacted with the prompted GPT that was configured with specific prompts derived from a guidance on the preliminary investigation.
• Figure 3: Box plot illustrating significant differences between Default GPT and Prompt GPT groups: participants in the Default GPT group showed greater disfavor toward the intruder and more favorable attitudes toward other individuals involved in the incident.
• Figure 4: The left and middle correlation plots indicate that in the default GPT group, participants who rated higher on self-perceived memory clarity are more likely to consider the GPT output accurate and reliable, and are more likely to judge the intruder guilty. The right correlation plot shows that in the guided prompt group, participants' actual memory recall scores align with their perceived clarity of memory regarding the observed incident. Two data points overlap in the Prompted GPT Condition for Perceived GPT Accuracy.
• Figure 5: An illustration of how participants co-write the statement with GPT, during which they modify the structure of their statement and also recall new details. Left: An illustration of the mental model of this co-operate interaction pattern. Right: An actual example from participants' chat history.