What if the kid who wins every school spelling bee still can’t explain how reading works? That is basically the mood of this paper. Imagine a child who can name every dinosaur and fruit pouch on sight, but does it with such weird private logic that no teacher would use that child as a model for how brains learn to see. Cute party trick, slightly alarming science problem.

For a while, neuroscientists had a very tempting story. As artificial vision systems got better at recognizing images, they also seemed to look more like primate vision under the hood. Early deep neural networks lined up surprisingly well with activity in parts of the ventral visual stream, especially inferotemporal cortex - the brain’s object-recognition department.

That overlap made people think AI progress and brain science might ride the same elevator upward. Better AI, better model of biology. Nice. Tidy. Suspiciously tidy.

The new Trends in Cognitive Sciences paper by Drew Linsley, Pinyuan Feng, and Thomas Serre argues that this alignment is flattening out, and in some cases getting worse, even as AI systems hit human or superhuman performance on vision benchmarks (Linsley et al., 2025). In other words, the machines may be acing the test while using the sort of reasoning that would make a biology professor put their head gently on the desk.

Same answer, different brain

Think of it this way: you and a GPS can both get to the coffee shop, but matching output is not the same thing as matching strategy.

That distinction matters because biology is not just about whether a system gets the answer right. It is about how it gets there. The paper’s core point is that modern vision models may solve tasks with shortcuts, training diets, or internal representations that drift away from how primates actually process visual information.

That concern is not coming out of nowhere. A 2022 Science Advances study found that the neat textbook story about deep network layers mapping cleanly onto the visual hierarchy breaks down when you compare models and brains more directly, especially once feedback signals enter the picture (Sexton & Love, 2022; PMCID: PMC9278854). A 2024 Nature Communications paper surveying 224 models found surprisingly tiny differences in brain predictivity across many architectures and objectives, while training data diversity mattered a lot more than people may have hoped (Conwell et al., 2024). Less "we found the brain recipe," more "apparently the pantry matters."

Why this matters outside an academic pillow fight

If your goal is just to label internet images, maybe none of this is a crisis. But if your goal is neuroscience - explaining how biological vision works - then benchmark worship can quietly send the field into a ditch.

This paper is really a warning about confusing engineering success with scientific explanation. A model can be useful and dazzling while still being a bad stand-in for living tissue. Jets do not flap, submarines do not grow gills, and a system trained on giant internet datasets does not become a faithful cousin of macaque visual cortex just because it is very good at spotting a toaster.

That matters for medicine and technology too. If researchers want better brain-computer interfaces, prosthetic vision, or tools for diagnosing visual disorders, they need models that capture the brain’s actual constraints, not just leaderboard swagger. Recent work suggests that adding biologically relevant pressures can help. Training CNNs on blurry images, for example, improved robustness and made them more human-like in several perceptual tests (Jang & Tong, 2024). That is the kind of move this paper is pushing for: not bigger by default, but truer by design.

So what should vision science do now?

The authors’ answer is refreshingly unromantic. Vision science should chart its own course. Build models grounded in biological data. Test them against neural activity, behavior, and real-world visual challenges. Stop assuming that whatever wins at internet-scale prediction has also cracked the code of primate perception.

That does not mean AI and neuroscience should break up and divide the furniture. AI can still be a powerful tool for neuroscience. Harvard’s Talia Konkle said in a 2024 interview that these models give researchers new tools for asking basic questions about vision. True. But tools are not explanations.

The fun part, if you like scientific messiness, is that this may push the field toward better questions. What visual experiences shape brain-like representations? Which architectural features actually matter? Where do humans rely on feedback, context, embodiment, and development in ways current AI mostly hand-waves past?

So yes, smarter AI is impressive. But this paper’s point is sharper than that: if you want to understand biology, you cannot just keep making the machine louder and call it a bird.

References

1. Linsley D, Feng P, Serre T. Better artificial intelligence does not mean better models of biology. Trends in Cognitive Sciences. 2025. DOI: https://doi.org/10.1016/j.tics.2025.11.016
2. Sexton NJ, Love BC. Reassessing hierarchical correspondences between brain and deep networks through direct interface. Science Advances. 2022;8(28):eabm2219. DOI: https://doi.org/10.1126/sciadv.abm2219 PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC9278854/
3. Conwell C, Prince JS, Konkle T. A large-scale examination of inductive biases shaping high-level visual representation in brains and machines. Nature Communications. 2024;15:9383. DOI: https://doi.org/10.1038/s41467-024-53147-y
4. Jang H, Tong F. Improved modeling of human vision by incorporating robustness to blur in convolutional neural networks. Nature Communications. 2024;15:1989. DOI: https://doi.org/10.1038/s41467-024-45679-0
5. McGrath SW, Russin J, Pavlick E, Feiman R. How Can Deep Neural Networks Inform Theory in Psychological Science? Current Directions in Psychological Science. 2024;33(5):325-333. DOI: https://doi.org/10.1177/09637214241268098 PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC11824574/

Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.