AI Vision Has a Blind Spot: Big Objects Slip Out of View

Modern AI vision models are trained by simply showing them millions of images and asking them to predict what goes together — no human labels required. A new paper from researcher Mayank Singal examined what those models actually learn, and found something quietly important. The models do learn that two patches of pixels belong to the same object. But that skill has a distance limit. The farther apart two pieces of the same object are in an image, the less likely the AI is to treat them as one thing, and the linking fades in a smooth, predictable curve.

Think of it like a flashlight beam. The AI can tell that your car's front bumper and its side mirror belong together. But widen the view to a whole bus, or a sofa stretching across a living room photo, and the AI starts seeing separate pieces instead of one object. The study found the same curve across different models and image sets, which suggests it's a fundamental trait of how these systems learn, not a bug in one product.

That single finding explains several odd behaviors people have noticed. AI tends to handle small objects better than large ones. It more often mixes up two objects of the same type — say, two identical chairs pushed together — than a chair and a table. And it usually keeps an object's parts grouped with the whole, like a wheel with its bicycle. Surprisingly, being partly hidden behind something else didn't matter much once size was accounted for.

The practical takeaway: the headline accuracy scores you see for AI vision hide this weakness. This paper doesn't offer a fix, and it hasn't been through peer review yet. But for anyone building or buying AI that looks at images — warehouse robots, medical scans, security cameras, photo organizers — knowing where the AI's eyes go blurry is the first step toward trusting it less in exactly those moments.