Stare at something moving for a while, then look at something still, and the still object seems to sit in a slightly different place. Primate brains reproduce that shift. Today's AI vision models do not, researchers at York University in Toronto report in Current Biology.
The illusion is called the motion aftereffect. The image does not change at all, only the viewer's recent experience does. That makes it a clean test of whether a system sees the way we do or only gets the right answer.
People, monkeys and machines
The team, led by Kohitij Kar, a Canada Research Chair in visual neuroscience, showed moving and still images to 79 people, two macaque monkeys and nine artificial vision networks.
In the monkeys, activity in the inferior temporal cortex, a brain area that recognizes objects, shifted the way people's perception did. The AI networks did not show that shift, explains graduate student Elizaveta Yakubovskaya.
"Current AI vision systems can often determine where an object is accurately, but they generally do not reproduce the way recent visual experience can reshape that answer," Kar says.
Why copy a mistake
An illusion is a kind of error, so why would anyone want AI to make it? Kar's argument is that the error reveals how human vision works.
"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 says. On this test, the machines still see differently.
