A CHI 2025 experiment from the MIT Media Lab is drawing fresh attention for an uncomfortable finding: people who view AI-edited images or AI-generated videos are markedly more likely to misremember what they originally saw—and more confident that their false recollections are correct. The study, co-authored by memory researcher Elizabeth Loftus of the University of California, Irvine, tested 200 participants in the United States and has new relevance as smartphone makers ship one-tap generative editing tools.
- False memories roughly doubled: The group shown videos made from AI-edited images erred on 38.7% of questions about altered details, versus 18.9% for the control group.
- Confidence rose with error: Participants in the worst-performing group rated their wrong answers 5.41 out of 7, compared with 4.54 in the control group.
- Labels did not stop the effect: Every condition saw an “enhanced image” label, yet false memories still formed—and skepticism or familiarity with AI filters offered no significant protection.
What the experiment actually measured
The protocol was tightly controlled. Two hundred volunteers viewed 24 royalty-free photographs for two minutes—political figures, travel scenes, archival material. After a two-minute round of Pac-Man as a distraction, they saw a second pass. One group saw the original images again. A second saw stills altered with Photoshop tools that added or removed people, objects and environmental details. A third watched five-second AI-generated videos derived from the untouched originals. A fourth watched AI-generated videos made from the already edited images.
Participants then answered 24 questions, one per image, and rated their confidence. The headline result is stark, but the underlying scale is smaller than it first appears. The control group missed 2.27 of the 12 questions about altered details; the worst group missed 4.64. That is a little more than two extra errors—enough to matter in forensic, journalistic and personal contexts, but not evidence that AI turns memory into fiction on its own.
“Basically, I think everything is very dangerous,” Pattie Maes, an MIT Media Lab professor and co-author, said of the findings.
The paper also breaks the effect down by edit type. In the video-of-edited-image condition, people-related edits produced the largest absolute share of false memories, about 45%, up from 19%. Environmental edits showed the sharpest relative increase, rising from 9.5% to 33.3%, while object edits moved from 15.6% to 36.7%.
Is the problem AI, or manipulation?
That is the most important caveat in the study, and the authors do not fully resolve it. There was no group shown conventionally retouched photos, so the experiment cannot prove that generative AI distorts memory more than a skilled manual composite. What changed across conditions was whether an image had been altered, not whether a model or a human did the altering.
The results point toward content alteration as the active ingredient. Volunteers who watched AI-generated videos made from unedited photographs did not show a statistically significant effect compared with the control group. The synthetic origin alone was not enough; the changed scene was.
An independent study from Georgetown University and the University of Washington reinforces that distinction. When 331 participants were shown summaries of a video, labeling them as AI-generated or human-made did not appreciably change whether they accepted false details. What mattered was whether the summary was accurate or misleading: recall accuracy fell from 83.6% to 44.8%.
The MIT result also complicates the usual remedy. All participants saw an “enhanced image” label on the modified visuals, mirroring the AI tags now appearing on social platforms. False memories formed anyway. Skepticism, familiarity with AI filters, gender and education level showed no significant protective effect. Age had a small but significant relationship. Knowing that an image may be manipulated does not appear to stop the brain from absorbing the manipulated version.
What the study does not prove
The authors are candid about the limits. The memory test was immediate, so it says nothing about whether these errors persist for weeks or months. The experiment covered only images and video, not synthetic text or audio. Exposure lasted two minutes, far less than the repeated encounters of daily life. The sample was 200 people in one country, and the stimulus set was 24 images. No countermeasure—media-literacy training, automatic detection or provenance checks—was tested.
The practical signal is narrower than the alarmist headline. If alteration rather than synthetic origin drives the effect, then the most useful defenses are likely to be procedural: keep unedited originals, preserve signed metadata and treat provenance as part of the image rather than a caption added later. Standards such as C2PA aim at exactly that problem, though they are not a cure-all. Detection arrives after perception, and this experiment suggests that by the time a viewer doubts an image, memory may already have incorporated it. The more defensible conclusion is that AI editing raises the stakes for verification—not that it creates false memories from nothing.