- A September 17, 2026 study in iScience, led by Biyu J. He of New York University, tested over 200 deep neural networks and found they consistently underperformed humans at object recognition when only global shape cues were available.
- The NYU team isolated shape, texture and internal detail as separate variables, revealing that DNNs exploit local statistical patterns rather than the silhouette-level reasoning humans default to when images are degraded.
Over 200 state-of-the-art deep neural networks, tested against human observers on distorted images, all lost to the humans on the same task: identifying objects from their overall shape. The September 17, 2026 iScience paper, led by New York University’s Biyu J. He, argues this is not a data problem or a scale problem but a structural one, with machines and humans using fundamentally different visual strategies.
The Shape Gap
Human object recognition is remarkably tolerant of degradation. Scratch away the texture, occlude the internal details, leave only a silhouette, and most people still identify the object without effort. The brain treats global shape as a primary cue, not a fallback. DNNs work the other way around. Trained on massive image datasets, they become extremely good at associating local pixel patterns and textures with labels. That works well on clean images. It breaks down when you strip the image to its outline.
He’s team designed their experiment to make this visible. They took images of common objects, including cats, butterflies, cars and corn on the cob, and systematically altered three visual properties independently: global shape, internal parts and texture. Presenting each altered version to both human participants and the neural networks allowed the researchers to isolate exactly which cue each type of processor was relying on. The results were consistent across every DNN architecture tested: when recognition depended on global shape alone, the AI models fell short of human performance. That pattern held across more than 200 networks, ruling out a quirk in any single design.
Why Benchmarks Miss This
Standard computer vision benchmarks measure accuracy on clean, well-lit, correctly oriented images. On those tests, the best DNNs match or beat humans. The NYU study shows that metric doesn’t capture what happens when conditions shift. A model that scores well on ImageNet may still fail to recognise a partially occluded pedestrian or a car viewed through fog, because those scenarios force reliance on global form.
He’s framing is direct: current AI models do not accomplish visual object perception the same way humans do, according to the paper. The models appear to learn statistical correlations between local features and object categories, rather than building anything resembling an abstract representation of an object’s shape. That distinction matters beyond academic interest. Autonomous vehicles, surgical robotics and AI-assisted medical imaging all operate in environments where images are routinely degraded, occluded or viewed from unusual angles. A vision system whose recognition degrades sharply under those conditions is a reliability risk, not just a benchmark footnote.
Where Current Architectures Fall Short
The pervasiveness of the finding across 200-plus architectures is the detail that makes this hard to dismiss. This is not one model family with a known weakness. The researchers found a consistent pattern across the field, which points toward something in how DNNs are trained rather than how any individual network is built.
The leading hypothesis in the paper is that standard training on large passive datasets reinforces texture bias. Models see millions of images where texture and label co-occur reliably, so they learn texture as the primary signal. Shape, being harder to extract from pixel statistics alone, gets deprioritised. Benchmark pressure compounds this: optimising for accuracy on clean data gives no gradient signal for shape-based reasoning under distortion.
Possible Fixes
He’s team does not claim to have solved the problem, but the paper points toward several directions. One is architectural: networks that explicitly encode global shape, perhaps through biologically inspired pathways that separate “what” from “where” processing, rather than leaving shape as an emergent property of local feature maps. Another is in training data. Synthetic datasets and 3D object models could force a network to build shape representations before texture is added, reversing the usual order. Loss functions that specifically penalise failures on shape-only inputs are a third lever.
None of these is a quick fix, and the paper does not quantify how much any of them would close the gap. What the study establishes is the measurement framework: a systematic, controlled method for separating shape, texture and detail contributions to recognition accuracy. That benchmark gives future research something concrete to optimise against, which is more useful than the general claim that AI vision needs to improve.
The Generalisation Problem Behind the Gap
The shape finding is a specific instance of a wider brittleness that has been documented across AI vision for years. DNNs trained on one distribution of images often fail on inputs that look slightly different, a problem the field calls out-of-distribution generalisation. Humans handle this almost automatically: a car is a car whether it appears in a photograph, a sketch, a shadow or a fog-blurred frame. The iScience results add precision to why that gap exists. It is not just that DNNs haven’t seen enough data. Their internal representations are tied to the statistical regularities of their training sets in ways that human perception is not.
That is the harder problem He’s work surfaces. Scaling up training data or model size addresses the quantity of statistical patterns a model can memorise. It does not change the nature of what the model is learning. Getting from pattern matching to shape-based abstraction may require architectural changes, new training objectives, or insights from cognitive science that the field has not yet integrated at scale.
