AI Sees Colors That Aren't There

A black-and-white spinning disk can make people see colors that aren't there. Red, blue and green appear along its curved lines, then swap places when the disk reverses direction. The toy is called Benham's top, and it has puzzled researchers since the 1890s.

Now scientists in Japan have reproduced the illusion in an artificial neural network. The twist is in what happened when they changed what it had learned to see.

The researchers used PredNet, a system trained to predict a video's next frame by comparing its guess with what actually appears. After training on ordinary first-person footage, it produced faint colors when shown a rotating black-and-white Benham's top, even though the input held no color at all.

The team then simplified the training material to isolate the cause. A computer-generated town scene with a bright magenta pedestrian produced more vivid results. When they stripped the videos down to colored squares moving across a screen, the colors of those objects strongly shaped the colors the network generated.

Even a simple sequence of flickering lines could produce the effect. Rotation itself wasn't necessary. The timing of the changing black-and-white patterns mattered.

The findings suggest predictive learning may help explain why humans experience subjective color, a phenomenon studied for roughly two centuries. Our visual systems may learn associations between motion and color, then apply those expectations to ambiguous input.

The model isn't a replica of the human brain, and the researchers aren't claiming to have solved the illusion. Retinal processing likely still plays a role. But the experiment offers a new way to study how experience shapes what we see.

The next question is whether people raised in different visual environments develop different versions of the illusion. A spinning disk might be a surprisingly useful window into a lifetime of seeing.

Read the full story at Scientific Reports, June 16, 2026

Hot Take: Anyone who has asked ChatGPT for a book recommendation and received a title, author and ISBN for a book that doesn't exist could have saved the researchers some time.