An AI Predicted Which Artworks People Would Remember Better Than the Artists Themselves

An image is not memorable merely because it is beautiful, technically impressive or unusual. Something about its arrangement survives the extraordinary competition for attention in a person’s visual memory. Researchers have now tested two linked propositions: whether artists can intentionally design that survival, and whether an AI can foresee it better than the people who made the work.[1]

The October 2026 Psychological Science study by Trent Davis, Yifei Chen and Wilma Bainbridge began with an art contest. Participating artists were asked to create works intended to be remembered or forgotten. The researchers then assessed recognition in three experiments involving adults recruited through Amazon Mechanical Turk (119), Prolific (892) and a Chicago gallery (62). These numbers describe the reported participant pools, not a claim that every person experienced identical test conditions.[1]

Artists achieved some of what they intended. Their memorable pieces were remembered more successfully. But the effect was asymmetric: artists found it easier to increase memorability than to deliberately suppress it. A painting intended to recede from memory may still contain a visually distinctive configuration the maker did not recognize as memorable.[1]

An artist can create an effect without being its best judge

A neural network called ResMem was more consistent than artists or other human observers at predicting which artworks would later be recognized. This result reverses an appealing intuition. We may expect the creator to understand the emotional and perceptual power of a work. The experiment instead distinguishes two different capacities: producing an image with a desired property, and accurately predicting how unfamiliar viewers will encode it.[1]

The team also examined visual characteristics that were interpretable in psychological terms. Its data-driven features explained nearly half the variance in memory outcomes and aligned with the computational predictor’s judgments more closely than with human expectations. That is not the same as proving the model discovered a single universal formula for memorable art. It is evidence that useful signal exists in measurable features whose relationship to memory exceeds ordinary introspection.[1]

The distinction matters to anyone commissioning imagery. Marketing departments seek images that linger; museum designers may want visitors to recall an important exhibit; educators choose diagrams students should remember. But it is dangerous to leap from a laboratory recognition task to claims about long-term persuasion, purchasing behaviour, emotional attachment or educational achievement. Remembering that an image appeared is only one aspect of what images can do.[1]

Why forgetting may be harder to engineer

Increasing memory can exploit distinctive features that attract attention or create strong perceptual structure. Deliberate forgettability has a different challenge: removing one distinctive cue may leave another. Viewers also bring different prior experiences. An ordinary object for one person might be personally evocative for another. The study establishes the observed asymmetry, not one definitive explanation for it.[1]

This limitation is intellectually useful. Art has always invited competing theories about what makes an image meaningful. Memory studies approach a narrower, measurable question. A computational predictor can score a likely recognition outcome yet say nothing about whether a picture is moving, ethically appropriate, worth conserving or aesthetically important.[1]

There is also a problem of optimization. If every visual creator begins designing to maximize the same predictor’s score, audiences may encounter a familiar palette of memorable motifs. The images that score well are not necessarily the images that best serve their purpose. Conversely, a striking artwork that performs modestly on recognition may be invaluable for a reason the dataset never measured.[1]

A new instrument, not an artistic referee

The most plausible use of this research is as an extra instrument in the studio: compare alternatives, understand likely attention and memory effects, and then decide whether those effects are wanted. It would be a mistake to delegate the artistic decision itself to the score.[1]

The finding is notable because it avoids the lazy question of whether a machine is more creative than a person. In this experiment, people created the works; AI predicted a measurable consequence of viewing them. That division of labour may turn out to be more useful than asking either side to imitate the other.[1]

Report an error or suggest a correction