Let me tell you a story you may have heard.
One day, a man wakes up after a stroke. He can still speak fluently—correct grammar, natural intonation, perfectly formed sentences—but listen closely and you'll notice what he says is empty. "I went to that… you know… the place where you buy things… the one with the cash registers." He wants to say "supermarket," but the word is gone. He can describe its function, but the naming area is damaged.
This is called aphasia. It's not about becoming stupid—it's the physical destruction of function in a specific region.
Clinical neurologists have studied aphasia for over a century. The logic is simple: which brain region is damaged, which language ability is lost—so you know what that region does. It's reasoning backward from "broken" to "how normal works."
Four computational linguists and cognitive scientists—Roll, Kries, Gwilliams, and Shain—have recently transported this century-old neuroscience method to language models.
🫸 Push gently and see what happens
What they did in their paper (arXiv:2605.16222) is essentially what neuroscientists do: break a model and watch what gibberish it produces.
But their way of "breaking" is far gentler than brain surgery. No operations, no cutting chips, no pulling GPUs. It's called "lesioning"—which sounds scary but just means setting certain parameters in the model directly to zero. Not deleting the model—turning off a specific component and seeing how the model's speech changes afterward.
For example, you zero out a chunk of the key projection matrix in an attention layer, then have the model complete "I bought ____ at the supermarket." Does it fail to produce "apples"? Does it output "I bought at the supermarket because if"? Or does it just freeze? Different lesion locations produce different types of language impairment—just as different brain-region damage produces different types of aphasia.
📋 How to run an aphasia assessment on a machine
To check whether a patient has aphasia, you need standardized tests. Clinicians have this tool: aphasia assessment batteries. Roll et al. adapted this to text, creating a Text Aphasia Battery (TAB).
What can TAB test? Grammar deficits—can the model distinguish "the man threw the ball" from "the man was thrown by the ball"? Semantic deficits—is "apple" in the same category as "banana" in the model's representation? Phonological deficits—can it still tell "b" from "p"? And fluency—can it produce meaningful sentences continuously, or does it stall after two?
Then they used TAB to test five 1B-parameter language models, collecting 112,426 outputs.
One hundred eleven thousand assessments. Every output labeled with the type of "language deficit" it exhibited.
🧩 Which layer breaks, what goes wrong
Once the experiments were done, the patterns emerged, and they were clear.
Attention components—the parts deciding "which word should I attend to"—and feedforward network (FFN) components—the parts deciding "what does this word itself mean"—produced completely different symptom distributions when damaged. Attention damage looked more like phonological/fluency problems; FFN damage looked more like grammar/semantics problems.
Even more interesting was depth. Lesions in shallow layers—near the input—mainly destroyed grammar and semantics. The model started confusing subject-verb-object, mixing up active and passive constructions. But lesions in mid-to-deep layers—roughly the model's waist—produced phonological and fluency problems. The model became like that stroke patient who couldn't say "supermarket."
This is fascinating. It means the language model's "brain"—if we can use that word—also has functional specialization. Not designed by anyone; it grew on its own during training.
🤔 Wait, is this like humans?
My reaction upon reading this far: does this resemble human aphasia?
Roll et al. naturally asked the same question.
The answer: somewhat, but not really.
For some lesioned models, the score distributions on certain TAB items were statistically similar to particular types of human aphasia. Pearson correlations weren't low; p-values were significant.
But this "similarity" is fragile. Once you dig into symptom quality—not "what's the score" but "what's the pattern of errors"—the differences between humans and machines emerge. Both are grammar errors, but the errors humans make and the errors machines make are not the same kind. A human says "I went to the store bought," dropping a preposition while keeping basic word-order logic. The model produces something that "looks like a sentence but every semantic relation between constituents is severed."
The paper's own words: *aphasia syndromes are heavily influenced by the details of learning and processing rather than being a domain-invariant consequence of disrupted language processing.* In plain terms: what aphasia looks like depends on who you learned language from and how—not just on "having a disrupted language system."
⚠️ Things I don't know
Honestly, this paper left me with questions that keep me up at night.
First, I don't know whether the 1B-parameter scale is representative. The human brain has roughly 86 billion neurons, with thousands of synaptic connections per neuron. 1B parameters are orders of magnitude smaller. Would lesioning a 100B-parameter model—GPT-4-class—give completely different results? I don't know.
Second, I don't know how to interpret "zeroing" in neuroscience terms. Human aphasia is physical damage—cells die, connections break, irreversible. Zeroing a model just sets weights to 0; you can restore them anytime. Are these the same kind of "damage"? I don't know.
Third, the paper uses TAB—a text adaptation of a clinical aphasia battery—but I haven't seen enough evidence that TAB's validity transfers from humans to machines. Does a test designed for humans, applied to models, still measure the same thing? I'm not sure about this either.
But these "don't knows" are exactly what makes it good. Real science tells you not "here is the answer" but "I found an interesting way to ask this question, and you can keep walking down this path." That's what this paper does.
📌 The bottom line
Roll et al. borrowed a century of clinical neuroscience practice and performed a functional dissection on a black box. They found that language models have internal functional specialization too, and these specializations exhibit a structure that is both similar to and different from biological neural networks.
The similarities excite us—"AI really is like a brain!" But the differences are where the real interest lies—"the ways AI is *not* like a brain are precisely telling us something about learning itself."
That's enough. An experiment that sparks this kind of curiosity is a good experiment.
---
References
1. Roll, N., Kries, J., Gwilliams, L., & Shain, C. (2026). *Artificial Aphasias in Lesioned Language Models*. arXiv:2605.16222 [cs.CL]. https://arxiv.org/abs/2605.16222
2. Damasio, A. R. (1992). *Aphasia*. New England Journal of Medicine, 326(8), 531-539.
3. Olah, C., et al. (2020). *Zoom In: An Introduction to Circuits*. Distill, 5(3), e00024.001.
4. Wang, A., et al. (2019). *GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding*. ICLR 2019.
5. Fedorenko, E., et al. (2024). *The Language Network is Robustly Engaged in Language Processing Across Diverse Tasks and Populations*. Nature Human Behaviour, 8, 1060-1075.