Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection

TL;DR

This study demonstrates that AI-edited images and AI-generated videos markedly increase false memories, with the strongest effect when AI-edited images are transformed into AI-generated videos. Using a preregistered, between-subjects design with 200 participants across four conditions, the authors show that AI manipulation not only elevates the frequency of false memories but also boosts confidence in those memories, particularly for dynamic video content. The work highlights critical implications for memory, perception, and decision-making in daily life, media literacy, and legal contexts, while also exploring potential therapeutic and self-esteem applications under strict ethical safeguards. It further discusses labeling limitations, the need for proactive mitigation strategies, and avenues for future research across diverse content types, modalities, and populations. Overall, the findings underscore the substantial cognitive risks posed by AI-enabled media manipulation and call for interdisciplinary collaboration to design safeguards, inform policy, and guide ethical deployment of memory-modifying technologies.

Abstract

AI is increasingly used to enhance images and videos, both intentionally and unintentionally. As AI editing tools become more integrated into smartphones, users can modify or animate photos into realistic videos. This study examines the impact of AI-altered visuals on false memories--recollections of events that didn't occur or deviate from reality. In a pre-registered study, 200 participants were divided into four conditions of 50 each. Participants viewed original images, completed a filler task, then saw stimuli corresponding to their assigned condition: unedited images, AI-edited images, AI-generated videos, or AI-generated videos of AI-edited images. AI-edited visuals significantly increased false recollections, with AI-generated videos of AI-edited images having the strongest effect (2.05x compared to control). Confidence in false memories was also highest for this condition (1.19x compared to control). We discuss potential applications in HCI, such as therapeutic memory reframing, and challenges in ethical, legal, political, and societal domains.

Figures (10)

• Figure 1: The figure illustrates how AI-generated content can potentially create false memories, particularly through AI-altered videos on social media platforms like TikTok. A recent trend on these platforms involves using AI to animate photos of deceased relatives, creating simulated interactions. These artificial experiences may blur the line between genuine memories and digitally fabricated ones, potentially affecting how people remember their loved ones.
• Figure 2: This figure illustrates the study procedure for our experiment examining how AI-generated images and videos can induce false memories. Participants first viewed original images to establish baseline memories, then were exposed to AI-modified versions after a filler task. These modifications included changes like increased military presence or removed climate change indicators. Finally, participants' memories of the original images were assessed through a series of questions, allowing researchers to measure the impact of AI-edited visuals on recall accuracy.
• Figure 3: The stimulus set consisted of four distinct categories: unedited images, AI-edited images, AI-generated videos from unedited images, and AI-generated videos from AI-edited images. The edits were further divided into three subgroups based on the type of change: People, Objects, and Environment. In the questionnaire, masked versions of the images were used to facilitate recall without revealing the edited features.
• Figure 4: Survey interface components. From left to right: (1) Original image viewing instructions, (2) AI-enhanced image viewing instructions, (3) Questionnaire about original image details and confidence assessment. The sample images depict a wedding scene. The questionnaire prompts participants to recall specific details and rate their confidence in their memory.
• Figure 5: The percentage of reported false, uncertain, and non-false memories (i.e., how many times participants recalled incorrectly, were unsure, or recalled the original image correctly) were analyzed using a one-way Kruskal-Wallis and post hoc Dunn with FDR. P-value annotation legend: **, P<0.01; ***, P<0.001; ****, P<0.0001.