Let me tell you a story you may have heard before.
One day a person wakes up after a stroke and can still speak fluently—grammar intact, natural intonation, perfectly formed sentences—but listen closely and you'll notice everything they say is empty. "I went to that... you know... the place where you buy things... the place with the cash registers." They want to say *supermarket*, but the word is gone. They can describe the function, but the naming area is destroyed.
This is called aphasia. Not stupidity—a specific region's function has been physically damaged.
Clinical neurologists have studied aphasia for over a century. The logic is simple: whichever brain region is broken, whichever language ability is lost—you now know what that region handles. Reasoning backward from "broken" to "how normal works."
Four researchers in computational linguistics and cognitive science—Roll, Kries, Gwilliams, and Shain—recently ported this whole enterprise to language models.
Gently Pushing to See What Happens
What they did in their paper (arXiv:2605.16222) is essentially what neuroscientists do: break a model, then watch what nonsense it produces.
But their "breaking" is far gentler than brain surgery. No scalpels, no chips removed, no GPUs unplugged. They do something called lesioning—which sounds scary but just means setting certain parameters in the model directly to zero. Not deleting the model; switching off a specific component and seeing how the model's speech changes.
For example: zero out one block of the key projection matrix in an attention layer, then ask the model to complete "I bought ____ at the supermarket." Does it fail to fill in "apples"? Does it produce "I bought at the supermarket because if"? Or does it just stall? Different zeroed-out locations produce different types of language impairment—just as different brain lesions produce different types of aphasia.
How to Run an Aphasia Assessment on a Machine
To check whether a patient has aphasia, you need a standardized test. Clinicians have one: the aphasia assessment battery. Roll et al. moved this to text, building a tool called the Text Aphasia Battery (TAB).
What can TAB measure? Grammar deficits—can you distinguish "The man threw the ball" from "The man was thrown by the ball"? Semantic deficits—do "apple" and "banana" belong to the same category in your representation? Phonological deficits—can you still hear the difference between "b" and "p"? And fluency—can you produce continuous meaningful sentences, or do you stall after two?
They then used TAB to assess five 1B-parameter language models, 112,426 outputs in total.
A hundred and eleven thousand evaluations. Each output labeled with which type of "language impairment" it displayed.
Which Part Broken Causes What
When the experiments finished, the patterns emerged—and they were clear.
Attention components (the parts deciding which word the model "should attend to") and feedforward network components (the parts deciding "what this word itself means") produced completely different symptom distributions when damaged. Attention damage looked more like phonological-fluency problems; FFN damage looked more like grammar-and-semantics problems.
Even more interesting: depth. Shallow-layer damage—layers near the input—mainly destroyed grammar and semantics. The model started confusing subject-verb-object roles, mixing up active and passive constructions. But mid-to-deep-layer damage—around the model's waist—produced phonological and fluency problems. The model became like that stroke patient who couldn't produce "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 out of training on its own.
Wait—Is This Actually Like Humans?
My reaction at this point: does this actually resemble human aphasia?
Roll et al. asked that too.
The answer: somewhat, but not quite.
Some lesioned models did produce score distributions across TAB items statistically similar to certain human aphasia types. Pearson correlations weren't low; p-values were significant.
But this resemblance is fragile. Once you dig into symptom *quality*—not "what's the score" but "what does the error pattern look like"—human and machine diverge. Both may make grammar errors, but they aren't the same kind. Humans say "I went to the store and bought," dropping a preposition but preserving the basic logic of word order. Models produce something that "looks like a sentence but the semantic relations between all constituents are 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 language: what aphasia looks like depends on who you learned language from and how—not just on "language system got disrupted."
What I Don't Know
Honestly, this paper left me with questions that keep me up at night.
First, I don't know whether 1B parameters is representative. The human brain has roughly 86 billion neurons, each with thousands of synaptic connections. 1B parameters is several orders of magnitude smaller. If you ran this on a 100B-parameter model—at the GPT-4 scale—would the lesioning results look completely different? I don't know.
Second, I don't know how to interpret "zeroing out" 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 actually 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 holds for machine models the way it does for humans. A test designed for people, applied to models: does it measure the same thing? I'm not sure.
But these unknowns are precisely 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 exactly what this paper does.
Bottom Line
Roll et al. borrowed a century of clinical neuroscience experience and performed a functional dissection on a black box. They found language models have internal functional specialization too—specializations that show a structure both similar to and different from biological neural networks.
The similarities make us excited: "AI really is like a brain!" The differences are what's truly interesting: "the ways AI is *unlike* a brain are telling us something about learning itself."
That's enough. An experiment that provokes 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.