Human Vision Has Quirks That AI Can’t Match

Human vision possesses some peculiarities that artificial intelligence (AI) systems struggle to replicate. Our eyes have blind spots, produce after-images, and fall prey to optical illusions. However, these very imperfections may be adaptive features of biological vision rather than mere flaws. Recent research from York University indicates that enhancing AI to mirror these human perceptual quirks could yield significant benefits.

One such illusion is the "motion aftereffect," where staring at an object moving in a particular direction can leave a still object appearing slightly displaced in the opposite direction. This perceptual phenomenon is akin to the wobble one experiences after disembarking from a rocking boat onto solid ground. The York University researchers aimed to explore the differences between human and machine vision by testing 79 humans, two macaques, and nine AI vision networks on the motion aftereffect.

While primate vision systems exhibited a response similar to humans, AI visual systems did not, even when prompted appropriately.

Senior author and neuroscientist Kohitij Kar emphasized the importance of replicating these human perceptual mistakes in AI computational systems. He stated that AI systems should be designed to not only accurately map objects in physical space but also reflect the ways that the human visual experience might alter that perception. Although the study's findings were modest, the researchers suggest potential avenues for incorporating this effect into AI systems through training and feedback.

Kar argued that focusing solely on the accuracy of AI answers is insufficient for creating systems that align with human perception and behavior. Instead, understanding the computations underlying human perception and behavior is crucial for developing AI that works more harmoniously with humans. While the proposed training methods remain hypothetical and untested, the broader implications are substantial.

More human-like AI visual systems could serve as valuable tools for neuroscientific research and improve the interpretation of medical images, driver assistance interfaces, and augmented reality applications. However, the potential for AI models to predict human perceptual errors also raises concerns about manipulation and the possibility of designing better strategies for deceiving humans.

As AI systems continue to advance, the question arises whether replicating human-like vision in AI is a path of peril or promise.