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

Pat Pataranutaporn, Chayapatr Archiwaranguprok, Samantha Chan, Elizabeth Loftus, Pattie MaesCrimRxiv 2025 · DOI

Abstract

Artificial Intelligence is increasingly used to enhance and modify visual media, often through tools integrated directly into consumer smartphones. This study investigates the impact of AI-altered visuals on the formation of false memories—recollections of events that either never occurred or deviate significantly from reality. In a pre-registered study involving 200 participants, we compared the effects of four media conditions: unedited images, AI-edited images, AI-generated videos, and AI-generated videos of AI-edited images. Our findings reveal that AI-edited visuals significantly increase false recollections, with the combination of content manipulation and video animation producing the strongest effect (a 2.05x increase compared to control). Furthermore, participants reported higher confidence in these false memories. These results suggest profound implications for legal testimony, digital ethics, and potential therapeutic applications in memory reframing.

Introduction

The phenomenon of "false memories"—the psychological process of recollecting events differently from how they occurred—has been a cornerstone of cognitive research for decades. Traditionally, these distortions were induced through verbal suggestion or basic photo manipulation. However, the advent of generative AI has introduced a new variable: high-fidelity synthetic media.

Today, smartphone users can "re-light" scenes, remove background figures, or animate still photos into realistic videos with minimal effort. This study seeks to understand how this shift from static, manual manipulation to dynamic, AI-driven synthesis affects human memory. As AI makes it easier to create hyper-realistic visual "evidence" of a fabricated past, how vulnerable does our cognitive recollection become?

Methodology

The researchers conducted a pre-registered study with 200 participants, divided into four experimental groups of 50. The study utilized a classic memory distortion paradigm consisting of four distinct phases:

  1. Encoding: Participants were shown a series of original, unmanipulated images representing everyday scenes.
  2. Distraction: A filler task was administered to allow the initial memories to settle and to create a temporal gap between exposure and recall.
  3. Re-exposure (The Intervention): Participants were shown "reminders" of the original images. The nature of these reminders depended on the assigned group:
  4. Control Group: Viewed the original, unedited images.
  5. AI-Edited Images: Viewed static images where specific elements were altered using AI.
  6. AI-Generated Videos: Viewed the original images transformed into moving videos via AI animation.
  7. AI-Generated Videos of AI-Edited Images: Viewed videos that animated the previously manipulated versions of the images.
  8. Testing: Participants were tested on their recollection of the original events to determine if the intervention media had distorted their primary memories.

To ensure ecological validity, the researchers created a custom dataset of visual stimuli designed to mimic personal photos and videos. These were processed using contemporary generative AI tools to ensure they matched the quality of media a user might encounter on a modern mobile device.

Key Insights: The Mechanics of Distortion

The Vividness Effect

The study suggests a correlation between the sensory richness of media and its ability to overwrite memory. AI-generated videos provide more immersive sensory cues than static images, making it more difficult for the brain to reject the false information. The "aliveness" of a moving image appears to bypass certain cognitive filters that might otherwise flag a static edit as suspicious.

The Confidence Trap

Perhaps the most concerning finding is that AI-altered visuals do not merely cause confusion; they bolster certainty. Participants exposed to AI-manipulated media were not just more likely to remember the wrong details—they were significantly more confident that those false details were accurate. This "confidence boost" presents a major challenge for contexts where the perceived reliability of a witness is paramount.

Double Manipulation

The most potent catalyst for memory distortion was the combination of a content edit (changing the facts of the scene) and a format shift (animating that change into a video). This suggests that the process of "animating a lie" makes it significantly more persuasive to human cognitive systems than a static manipulation alone.

Results

The quantitative data highlights a significant shift in memory reliability when AI is introduced into the feedback loop:

  • Increased False Memories: The condition involving AI-generated videos of AI-edited images resulted in a 2.05x increase in false recollections compared to the control group.
  • Boosted Confidence: Participants in the strongest AI condition reported a 1.19x increase in confidence in their false memories compared to those in the control group.
  • Statistical Significance: Across all tested metrics, AI-edited visuals significantly outperformed traditional unedited visuals in their capacity to distort recollection.

Discussion and Future Work

While the technical community remains focused on the detection of deepfakes through software, this research highlights a different vulnerability: the human "wetware." The brain is often easier to fool than a digital detector, and the psychological effects of synthetic media can persist even if the media itself is later identified as fake.

The findings raise urgent questions regarding the use of AI-enhanced media in legal and investigative contexts. If a witness is shown an AI-upscaled or "clarified" video of a crime scene, there is a high risk that the technology could inadvertently rewrite the witness's actual memory of the event, leading to confident but false testimony.

Therapeutic Potential

Conversely, the authors note a potential positive application in Human-Computer Interaction (HCI): memory reframing. For individuals suffering from PTSD or traumatic memories, AI-mediated visual tools could potentially help "re-author" painful recollections into more manageable forms. However, this application remains controversial and would require rigorous ethical frameworks to prevent misuse.

Limitations

Future research is required to determine the long-term persistence of these synthetic memories. While this study measured distortion shortly after exposure, it is currently unknown whether these false recollections remain stable over weeks, months, or years.

Summary updated

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.
BibTeX
@inproceedings{pataranutaporn2025synthetic,
  author = {Pat Pataranutaporn and Chayapatr Archiwaranguprok and Samantha Chan and Elizabeth Loftus and Pattie Maes},
  title = {{Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection}},
  booktitle = {CrimRxiv},
  year = {2025},
  doi = {10.21428/cb6ab371.22d2375b},
}

Loading related content...

Figure