Samantha Chan* (MIT Media Lab), Pat Pataranutaporn* (MIT Media Lab), Aditya Suri* (MIT Media Lab), Wazeer Zulfikar (MIT Media Lab), Pattie Maes (MIT Media Lab), and Elizabeth Loftus (University of California, Irvine)
*Equal contributions
Samantha Chan ([email protected]) & Pat Pataranutaporn ([email protected])
This paper investigates AI's impact on false memories --- recollections of events that did not occur or deviate from actual occurrences. The study explores false memory induction through suggestive questioning in Human-AI interactions, simulating crime witness interviews by AI systems. Four experimental conditions were used: a control, a survey-based condition, a pre-scripted chatbot condition, and a generative chatbot condition using a large language model (LLM). Participants (N=200) were randomly assigned to conditions in a two-phase study. In Phase 1, they watched a crime scene video, then interacted with their assigned AI interviewer or survey, answering questions about the video with five misleading ones. False memories were assessed immediately after. Phase 2, conducted a week later, evaluated false memory persistence. Results showed the generative chatbot condition led to significantly higher false memory formation rates. It induced over 3 times more immediate false memories than the control and nearly 1.7 times more than the survey-based method. The study also explored moderating factors influencing false memory formation. Findings highlight the potential risks of using advanced AI systems in sensitive contexts like police interviews and emphasize the need for further research and ethical considerations.
├── Data/
│ ├── Raw/
│ ├── Processed/
│ └── Code/
├── Prototype/
│ ├── Survey/
│ ├── Pre-Scripted_Chatbot/
│ └── Generative_Chatbot/
└── Supplementary/
├── Survey/
└── Video/
- Raw: Original, unprocessed, and de-identified data collected during the study.
- Processed: Cleaned and formatted data used for analysis.
- Code: Scripts and notebooks used for data analysis and visualization.
- Survey: Materials for the survey-based condition.
- Static Chatbot: Implementation of the pre-scripted chatbot.
- Generative Chatbot: Implementation of the LLM-based generative chatbot.
- Survey: Survey materials and questionnaires used in the study.
- Video: Crime-related video (2:30) from the Sayford Supermarket robbery on April 6, 2019, which was used in the experiment. Original video (4:17): https://www.youtube.com/watch?v=KEITMPG321Y&ab_channel=PennLive.com
