AI vision models do not fall for an illusion that fools monkeys and people

Stare at something moving in one direction, then look at a still object, and the object seems to sit slightly off to the side. York University researchers found that this shift shows up in the visual cortex of macaques as well as in human reports, but not in any of the artificial vision networks they tested, according to a study published October 2 in Current Biology.

A useful kind of mistake

The effect is called the motion aftereffect. Nothing in the second image moves, yet the viewer's sense of where things are has changed. That gap between what the pixels say and what is perceived makes the illusion a clean test: a system that reads position straight from the image should not be fooled, while one that works like biological vision should be.

The team paired experiments with human volunteers with recordings from the inferior temporal (IT) cortex of macaques, a brain region involved in recognizing objects. People reported the illusion, and the position signals in the monkeys' IT cortex shifted the same way, although the picture on the screen never changed.

Accurate, but not human-like

The researchers then ran the same test on three families of artificial networks: standard feedforward models, recurrent models and recent video-based models. All of them located the object correctly. None of them showed the shift that adaptation causes in people and monkeys, according to the paper, which was first posted as a preprint in March.

That makes the illusion a new benchmark for vision models that are supposed to process the world over time rather than one frame at a time. The authors are York graduate student Elizaveta Yakubovskaya, Hamidreza Ramezanpour, Matteo Dunnhofer and senior author Kohitij Kar, who holds the Canada Research Chair in Visual Neuroscience.

Should AI copy our errors?

Kar's argument is that matching human answers is not the same as matching human vision. "If we want AI that works with humans and understands the world in more human-compatible ways, we cannot focus only on whether it gets the right answer," he said in the university's announcement.

The odd conclusion is that an AI built to see like us might need to share some of our systematic mistakes. The study does not show that adding this kind of history dependence would make AI better at real tasks. That is the open question it leaves for model builders.