Object Recognition: Surprising Test Warning for AI Vision

Object recognition is the task computer vision is supposed to have solved, and a study published in iScience on 17 September 2026 says it has not. Researchers Mugihiko Kato and Biyu J. He of New York University built an image set that deliberately pulls apart the three cues a viewer can use to identify a thing — its overall outline, the parts inside it, and its surface texture — then ran more than 200 deep neural networks against human viewers on the same pictures.

No model reproduced the human pattern. That is the headline, and it is stated flatly in the paper: “No DNNs replicated humans’ cue-reliance profile, including those with recurrence or specialized training.” The sharper result sits one sentence later. When the overall shape of an object was the only cue left intact, every network tested underperformed people, without exception.

“Current AI models do not accomplish visual object perception in the same way that humans do,” He told reporters. “We tested more than 200 deep neural networks, and no models fully reproduced humans’ object recognition patterns.” That is a strong claim, and it is not the claim that computer vision does not work. It plainly does. The argument is that it works by a different route from ours, and that the difference surfaces precisely where reliable machine vision matters most.

This article covers the design of the test, what the numbers showed, which model families came closest, the result about brain-alignment scores that should worry anyone using them as a proxy, and the practical read for self-driving, robotics and assistive devices.

Table of contents

What the Object Recognition Study Actually Tested
Where Machine Object Recognition Falls Short on Global Shape
The Object Recognition Models That Came Closest, and Where They Broke
Brain Alignment Does Not Predict Object Recognition Behaviour
Why Global Shape Is Hard for Machine Object Recognition
The Longer Argument This Object Recognition Result Joins
What This Means for Systems That Depend on Object Recognition
What the Object Recognition Study Does Not Show
How to Read an Object Recognition Result Without Overreaching
Object Recognition: Frequently Asked Questions
References

What the Object Recognition Study Actually Tested

The design is the interesting part, because it is what lets the authors say something about how recognition happens rather than merely how often it succeeds.

The core idea

Everyday images give you every cue at once. A photograph of a cat has a cat-shaped outline, cat-shaped internal parts, and fur texture, so a neural network can score well while relying on any one of them. The study’s image set systematically untangles global shape, internal parts and texture information so that each object recognition cue can be isolated.

The object categories

The team worked with familiar, easily named things — a cat, a butterfly, a car, corn on the cob — and then distorted them along one axis at a time. Familiarity matters: it removes “the human did not know what that was” as an explanation for any gap.

The comparison group

Human object recognition behaviour was measured on the same perturbed images, giving a per-condition profile rather than a single accuracy score. That profile is what the models were then scored against.

The model sweep

More than 200 deep neural networks were tested, spanning diverse architectures, training diets and training objectives. That breadth is what makes the negative object recognition result meaningful: it is not a claim about one model but about a whole population of them.

The measure that matters

The paper does not simply ask which system is most accurate. It asks which system has the same cue-reliance profile as a person — meaning it succeeds and fails on the same conditions for the same reasons. That is a much stricter test of object recognition than a benchmark score.

Cue isolated	What the viewer still sees	Human performance	Model performance
Global shape only	Outline and silhouette	Strongest condition	All models substantially below humans
Internal parts	Component structure inside the form	Reliable	Closer, but profile still diverges
Texture	Surface pattern and material	Used, not relied upon	Relatively favoured by models
All cues intact	An ordinary photograph	Near ceiling	Near ceiling — the gap is hidden

Where Machine Object Recognition Falls Short on Global Shape

The paper’s strongest sentence is about a single condition, and it uses the word “strikingly.”

The global shape result

“Strikingly, all DNNs substantially underperformed humans when the global shape cue alone was critical to object recognition.” Every object recognition architecture, every training diet, every objective. Not a weak average — a uniform shortfall.

What that means in plain terms

If you erase the texture and scramble the internal detail of a butterfly but leave its outline, a person still says butterfly. The object recognition models tested do not, reliably. He’s summary was that humans’ ability to leverage the global shape cue “remains unparalleled.”

Recurrence did not rescue it

Networks with recurrent connections — often proposed as the architectural ingredient missing from feedforward object recognition models — did not replicate the human profile either. Neither did models given specialised training regimes aimed at exactly this gap.

Why this is not a data problem

The obvious objection is that the models simply need more pictures. He addressed it directly: “These results show that current computer vision models are not quite human-aligned, despite being trained on a massive number of pictures that humans have taken.” The object recognition training set is enormous and human-generated, and the gap persists anyway.

The generalisation failure underneath

The broader pattern is that machine object recognition degrades badly on incomplete or broken visual information. Partly hidden, erased or fragmented objects are where models misclassify or give up, and those are exactly the conditions a camera in the real world produces constantly.

The Object Recognition Models That Came Closest, and Where They Broke

Not every result was negative, and the shape of the partial success is informative.

Text-image contrastive models led

Fine-tuned text-image contrastive-trained models — the CLIP-style family that learns from paired pictures and captions rather than from labels alone — were the most human-like overall, and the finding held regardless of architecture. That is a statement about training objective beating architecture.

But they broke in the same place

“Fine-tuned text-image contrastive-trained models, regardless of architecture, were most human-like overall, but lost their human-alignment when the global shape was disrupted.” The best family collapses at precisely the condition the paper identifies as the core gap.

Why the caption signal helps

Captions describe objects at the level a person would name them. Training against language plausibly pushes a model towards a more category-level, less texture-bound representation — which is why this family leads. It just does not go far enough.

Architecture turned out not to be the object recognition lever

Convolutional networks and transformers both appear across the sweep, and the grouping that predicted human-likeness was training objective rather than network type. Anyone hoping a new backbone fixes object recognition alignment should read that carefully.

What is still missing

The paper does not claim to know what would close the gap. It establishes that the object recognition gap is systematic and persistent across a 200-model population, which is a precondition for anyone trying to close it.

What the study population covered, by count

Deep neural networks tested 200+

Models that replicated the human cue profile 0

Models matching humans on global shape alone 0

Cue dimensions untangled by the image set 3

Object categories named in coverage 4

Brain Alignment Does Not Predict Object Recognition Behaviour

This is the finding with the widest blast radius, and it has had the least coverage.

The claim

“Alignment with ventral stream neural recordings in an existing database did not predict alignment to human behavior.” Scoring well against recorded activity in the brain’s object-recognition pathway told you nothing about whether a model would behave like a person on these images.

Why that is awkward

A substantial research programme scores object recognition models by how well their internal activations predict neural recordings, on the reasonable assumption that a more brain-like model is a more human-like model. This result says the two measures come apart.

The companion result

“Model performance does not always predict its human-alignment.” Benchmark accuracy is not a proxy for human-like object recognition either. So the two most commonly used proxies both fail on this test set.

What that leaves

Behavioural comparison on carefully controlled stimuli — which is expensive, slow and exactly what this paper did. There is no cheap object recognition substitute currently validated.

The honest caveat

This is one image set, one behavioural paradigm and one neural database. A single decoupling result does not retire the brain-alignment programme; it does mean a brain-alignment score should not be quoted as evidence of human-like object recognition without further work.

Why Global Shape Is Hard for Machine Object Recognition

The result is less surprising once you look at how object recognition systems are built, and the prior literature points the same way.

Texture bias is a known object recognition failure mode

Convolutional networks have been shown for years to lean on texture more heavily than people do, and to classify with limited use of global object shape. This study extends that line with a much larger model sweep and a cleaner cue separation.

Shape is a long-range property

Texture is local: a patch is enough. Outline is global — it depends on relationships between pixels far apart in the image. Architectures that build up from local filters are structurally better suited to the first than the second.

Humans are doing something else

Human object recognition appears to construct a global form early and fill in detail afterwards. Whatever the mechanism, the behavioural signature in this study is that people hold on to identity when only the outline survives, and the models do not.

Occlusion is the everyday version

A pedestrian behind a parked car is a global-shape problem. So is a road sign half-covered by a branch. The conditions the paper constructs in the lab are the ordinary conditions of outdoor object recognition.

Why more data may not help

If the shortfall were a coverage problem, the largest-trained models in the sweep would be closest to humans. The reported pattern is that training objective, not scale, predicted alignment — which points at a representational difference rather than a data gap.

Proxy commonly used	What it is meant to indicate	What this study found
Benchmark accuracy	General visual competence	Does not always predict human-alignment
Ventral stream neural fit	Brain-like processing	Did not predict behavioural alignment
Recurrent architecture	Closer to biological vision	Did not replicate the human cue profile
Specialised shape training	Corrects known texture bias	Did not replicate the human cue profile
Text-image contrastive training	Category-level representation	Most human-like overall — until shape is disrupted

The Longer Argument This Object Recognition Result Joins

The paper does not arrive in a vacuum. It lands on top of roughly a decade of work asking the same question with smaller model sets, and its contribution is scale and control rather than novelty of the question.

The shape-versus-texture line of work

Earlier studies established that convolutional networks classify with far less reliance on global object shape than people do, and that they can be steered towards shape by changing the training distribution. Those results were persuasive but narrow — a handful of architectures, one or two perturbation types.

What 200 models buys you

A sweep that wide converts a claim about particular networks into a claim about a design space. When every point in the space fails the same condition, the natural conclusion is that the condition is not being addressed by anything currently in the space.

The recurrence hypothesis, tested and unsupported

One long-standing proposal is that human object recognition depends on feedback from later processing stages back to earlier ones, and that purely feedforward networks therefore cannot match it. This study included recurrent models, and they did not replicate the human profile either — which does not falsify the hypothesis but removes it as an easy answer.

The training-objective result is the constructive one

The single positive signal in the paper is that text-image contrastive training moved models closest to people. That is actionable: it points at supervision signal rather than at network depth or width as the axis worth exploring, and it is consistent with the intuition that naming things at the level people name them shapes the representation towards human categories.

Why the failure survives the best case

The awkward detail is that the leading family lost its advantage exactly where the gap is widest. A partial fix that evaporates under the hardest condition is a clue about mechanism: whatever contrastive training is adding, it is not a global-form representation robust to perturbation.

What a closing result would look like

A model that matched the human cue-reliance profile across all four conditions — including global shape alone — would be the first genuine evidence that machine object recognition had converged on the human strategy rather than a parallel one that happens to score well on clean photographs.

What This Means for Systems That Depend on Object Recognition

The applied stakes are the reason an object recognition result from a vision lab got general news coverage.

Autonomous driving

A car’s perception stack is object recognition under continuous partial occlusion, unusual lighting and unfamiliar viewpoints. A systematic weakness on degraded global shape is not an edge case there; it is the operating condition.

Robotics and manipulation

A robot that must grasp a partly hidden item has to infer the whole from a fragment. This study says that inference is where machine vision is furthest from human capability, which is a direct constraint on unstructured-environment robotics.

Prosthetics and assistive devices

The authors frame the upside in exactly these terms: better models of human object recognition could improve computer vision and support assistive technologies, including brain-computer interfaces intended to help people with disabilities see and act in the real world.

Safety-critical inspection

Any deployment where a missed object is expensive — medical imaging, industrial inspection, security screening — inherits the same profile. Benchmark accuracy on clean images will not surface it.

The procurement lesson

If you are buying an object recognition system, ask for performance on degraded and occluded inputs specifically, not for a headline accuracy figure. The whole point of this paper is that the two numbers can diverge, and the divergence is where risk lives. The same discipline applies to any computer vision pilot moving from demo to production.

What the Object Recognition Study Does Not Show

Overclaiming is the standard failure mode with results like this, so it is worth being precise about the limits.

It does not say object recognition fails

Models remain near ceiling on ordinary undistorted photographs. The study is about a specific and controlled divergence in cue reliance, not a general indictment.

It does not rank commercial products

The sweep is of research models spanning architectures and training regimes, not a comparison of deployed object recognition APIs. No vendor is named or scored.

It does not identify a fix

The paper establishes that the gap is systematic and persistent. It does not propose an architecture or training objective that closes it, and no such proposal is implied.

It does not settle the brain-model debate

The decoupling of neural fit and behavioural alignment is one result on one database. It is a caution about a proxy, not a refutation of the approach.

It is one image set

Generalisation across stimulus sets is its own research question. The finding is strong within the design tested, and replication on independent image sets is the natural next step.

How to Read an Object Recognition Result Without Overreaching

A short checklist for anyone summarising an object recognition paper for a non-specialist audience.

Check what was held constant

The value here is the cue separation. A study that simply reports “AI scores lower than humans” on natural images tells you far less, because you cannot say what drove the gap.

Check the object recognition population size

A negative result across 200-plus models is qualitatively different from a negative result on three. Breadth is what converts “these models failed” into “this class of model fails.”

Check whether the best case was tested

The strongest finding here survives the best-performing family — contrastive models led and still lost alignment on global shape. A result that only holds for weak baselines is much less interesting.

Separate the object recognition mechanism claim from the capability claim

“Machines recognise objects differently” and “machines recognise objects worse” are different statements. This paper supports the first everywhere and the second specifically in the global-shape condition.

Look for the applied hook the authors themselves state

Kato and He point at computer vision improvements and brain-computer interfaces. That is a narrower and more credible claim than the self-driving framing that coverage tends to add, and it is the one with the authors’ name on it.

Object Recognition: Frequently Asked Questions

What did the study find?

That more than 200 deep neural networks all failed to reproduce the human pattern of visual cue use, and that every one of them substantially underperformed people when the global shape of an object was the only usable cue.

Who ran it and where was it published?

Mugihiko Kato and Biyu J. He of New York University, published in iScience on 17 September 2026 under the title “Systematic image perturbations reveal persistent gaps between human and machine vision” (DOI 10.1016/j.isci.2026.117370).

How were the images changed?

The team built an image set that systematically untangles global shape, internal parts and texture, then distorted familiar objects — a cat, a butterfly, a car, corn on the cob — along one dimension at a time.

Which models did best?

Fine-tuned text-image contrastive-trained models, irrespective of architecture, were the most human-like overall. They lost that alignment when global shape was disrupted.

Does a more brain-like model behave more like a human?

Not according to this study. Alignment with ventral stream neural recordings did not predict behavioural alignment, and raw model performance did not always predict it either.

Does this mean self-driving cars are unsafe?

The paper does not test any deployed system. It does identify a weakness — recognising objects from degraded global form — that maps onto occlusion, which is a routine condition for vehicle perception.

What did the researchers say the work is for?

Improving computer vision and supporting assistive technologies, including brain-computer interfaces that could help people with disabilities better see and act in the real world.