When AI Remembers for Us

Photographs are often treated as evidence of the past, even though they can be misremembered, cropped or taken out of context. Generative tools add another possibility: making a scene that never happened look plausible, or changing details in a real one. The question is not only what we see, but what we believe we remember after seeing it.

In a preregistered study published at the CHI 2025 conference, 200 participants were divided into four groups. After viewing original images and completing an intervening task, participants received different stimuli: unedited images, AI-edited images, AI-generated videos, or videos generated from edited images.

Researchers then tested participants’ recollections. AI-altered visuals increased false recollections; the largest effect appeared in the video condition based on edited images, with a false-memory rate 2.05 times that of the control group. Confidence in false memories was also higher in that condition.

That figure describes a comparison between groups in one experiment. It does not mean that every person had twice as many false memories in everyday life, nor does it show that a single piece of content can permanently rewrite autobiographical memory.

The finding matters because memory does not work like an archive storing an immutable recording. People reconstruct events from what they experienced, saw and learned later. A plausible image can become part of that reconstruction.

The study does not justify the conclusion that provenance labels are useless; that was not the question tested. It suggests instead that visual manipulation can affect recollection, and that the effect deserves further study in settings closer to everyday use, over longer periods and with different kinds of content.

For journalism, family archives, museums and platforms, a file’s provenance becomes part of its history. Metadata and content credentials can document origin and some edits when they remain attached throughout the chain. On their own, they do not guarantee that a depicted scene is true, complete or correctly interpreted.

The same distinction applies to organisations that keep product photographs, video or records of production processes. A reliable archive requires technical controls as well as editorial practice: who created the content, what was changed and what use is appropriate.

If synthetic images can enter the way we remember an event, what evidence will we need to preserve in order to reconstruct a content’s history? And how will we distinguish a personal memory from an image encountered later?

The study involved 200 participants across four controlled conditions. Its findings demonstrate an effect under the experimental conditions described, not a universal estimate of risk or a clinical diagnosis.