AI-altered photos increased false memories and confidence in an experiment, researchers say

Seeing the original photograph first did not protect participants from later remembering it incorrectly. In a 200-person experiment, those shown AI-altered versions misremembered the originals more often—and expressed greater confidence in those false recollections. Gizmodo described the findings on October 18, following The Washington Post’s reporting on research by Pattie Maes and colleagues.

The same scenes, with different details

Maes, a professor at the MIT Media Lab, and her colleagues began by showing participants 24 photographs. The scenes were ordinary: a wedding, a city street and a race. The Washington Post described the experiment’s setup in an October 15 article by Ariana Eunjung Cha.

After a break, some participants saw the original photographs again. Others saw versions changed with AI. Both groups therefore returned to pictures they had encountered earlier, but only one group received the altered material before the later memory test.

The edits changed both the people pictured and the contents of the scenes. The Post described three kinds of alterations:

• Changing facial expressions.
• Inserting military personnel into photographs.
• Adding signs of climate change.

How large was the difference between groups? Gizmodo supplies no numerical effect size for either memory errors or confidence. The findings establish a reported direction of change, but do not quantify how much the altered photographs shifted either measure.

Gizmodo also cites a study from the previous year in which deepfakes and AI-generated text planted false memories about fabricated political scandals and reduced voting intention. That study concerned invented political events, rather than recall of previously viewed pictures. Its reported consequence extended beyond what participants remembered to whether they intended to vote.