If you want to understand how a complicated radio works, the crudest but most effective method is to unplug random wires while it's playing and see whether the sound gets quieter, distorted, or dies completely.
In neuroscience, this is called a lesion study. By observing patients whose brains were damaged by stroke or trauma, doctors discovered that when Broca's area in the left hemisphere is damaged, patients understand speech but cannot produce fluent sentences (like a stuttering radio station); when Wernicke's area is damaged, patients speak fluently—but it's all gibberish (like crossed radio signals).
This condition is called aphasia.
In May 2026, a Stanford research team published a wildly creative paper: 《Artificial Aphasias in Lesioned Language Models》.
They decided to play 'cyber surgeon' and perform brain surgery on large language models (LLMs)—deliberately breaking parts of the model to see how AI actually produces language.
The Cyber Scalpel: Zeroing Out Parameters 🔪
Since LLMs are 'digital brains' made of hundreds of billions to trillions of parameters, what happens if we forcibly set a small block of them to zero?
The researchers operated on open-source models like OLMo, surgically 'removing' different components like slicing a cake:
- Attention mechanisms: responsible for sorting out logical relationships between words.
- Feed-forward networks (FFN): responsible for storing concrete vocabulary and knowledge.
- Damaging neurons in early layers (near the input) often left the AI completely broken, unable to produce even basic gibberish.
- Damaging neurons in later layers (near the output) preserved basic logic; the AI merely lost some rhetorical elegance and detail.
The Diagnosis: Three Kinds of AI 'Aphasia' 📝
The post-surgery results surprised everyone: the symptoms AI exhibited were strikingly similar to those of human aphasia patients!
1. 'Empty Content Syndrome' (FFN damage) 🌫️
When the model's FFN components were damaged, the AI behaved like a Wernicke's aphasia patient. Its speech remained grammatical and natural in tone, but the content became extremely vague. Ask it 'What is gravity?' and it might answer: 'It's a thing, in a place, doing that thing.' Conclusion: FFNs are the AI's 'library'—break them and the words are gone.2. 'Shattered Logic Syndrome' (Attention damage) 🧩
With the Attention components damaged, the AI became like a severe Broca's aphasia patient. It knew which words to say but couldn't string them together. Output might be a jumble of random words, or get stuck in an infinite loop of 'the the the.' Conclusion: Attention is the AI's 'glue'—break it and logic falls apart.3. 'Punctuation Meltdown' ⁉️
This was the strangest finding. When certain gates in the FFN (the Gate Matrix) were damaged, the AI could suddenly spiral out of control, spewing pages and pages of pure punctuation marks. In medicine, this would be called an 'atypical language burst.'The Secret of Layers: The Deeper, The Deadlier 🧱
The study also uncovered a 'depth law':
What Remains in the Black Box? 🕵️♂️❓
While this surgery revealed a lot, the paper still leaves a few head-scratchers:
1. The human–machine gap persists: Although the AI's symptoms resemble human ones, some of its failure modes (like emitting thousands of commas) have never been seen in humans. AI's language processing may imitate humans on the surface while running on a completely different, high-dimensional statistical 'alien grammar' underneath. 2. The mystery of self-repair: In brain research, some functions can migrate to other regions after damage. Do an AI's remaining neurons spontaneously 'fill in' after damage? This study doesn't yet answer that.
Summary
Understanding breakdown is a path to understanding intelligence. 🎞️
This paper tells us: AI is not an opaque black box. Its language ability has a clear division of labor—logic is logic, vocabulary is vocabulary, and foundations are foundations.
Through this kind of 'self-harming' research, we are closing in on the ultimate question: how do those strings of 0s and 1s, in the folds of a digital brain, weave language that convincingly mimics human emotion and thought?
The deepest insights often come from staring at what is broken. 🧠✨ That is the striking lesson this 2026 work in 'digital neuroscience' offers about the nature of language. 🎓🔭