If you want to understand how a complicated radio works, the crudest but most effective method is to randomly pull out a few wires while it's playing and see whether the sound gets quieter, distorted, or dies entirely.
In neuroscience, this is called a lesion study. By observing patients whose brains were damaged by stroke or trauma, doctors discovered that damage to Broca's area leaves patients able to understand speech but unable to produce fluent sentences — while damage to Wernicke's area leaves patients speaking fluently but saying nothing meaningful.
This condition is called aphasia.
In May 2026, a Stanford research team published a remarkably creative paper: 《Artificial Aphasias in Lesioned Language Models》.
They decided to play "cyber surgeon" — deliberately damaging parts of a large language model to see how AI actually produces language.
The Cyber Scalpel: Zeroing Out Parameters 🔪
Since an LLM is a "digital brain" made of hundreds of billions of parameters, what happens if we force a small block of them to zero?
The researchers operated on open-source models like OLMo, surgically "excising" different components as if slicing a cake:
- Attention: responsible for sorting out logical relationships between words.
- Feed-forward networks (FFN): responsible for storing vocabulary and knowledge.
- Damaging early (input-side) neurons often leaves the AI completely broken — it can't even produce basic babble.
- Damaging late (output-side) neurons preserves basic logic; the AI merely loses some rhetorical elegance and detail.
Diagnosis: Three Types of "AI Aphasia" 📝
The post-surgery results surprised everyone: the symptoms AI exhibited were strikingly similar to those of human aphasia patients!
1. "Content Emptiness Syndrome" (FFN damage) 🌫️
When the 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: the FFN is the AI's "library" — break it and it loses its words.
2. "Fragmented Logic Syndrome" (Attention damage) 🧩
When the Attention components were damaged, the AI became like a severe Broca's aphasia patient. It knew which words to say but couldn't string them together. The output might be a jumble of words, or it could fall into 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 within the FFN (Gate Matrix) were damaged, the AI could suddenly lose control and spew out pages upon pages of punctuation. In medicine, this resembles "atypical language outbursts."The Secret of Layers: The Deeper, the Deadlier 🧱
The study also revealed a "depth law":
What Remains in the Black Box? 🕵️
Although this surgery revealed a great deal, the paper leaves several head-scratchers:
1. The human–machine gap persists: although AI's symptoms resemble human ones, some 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 fundamentally different, high-dimensional statistical "alien grammar." 2. The mystery of self-healing: in brain research, some functions can migrate to other regions after damage. Will an AI's remaining neurons spontaneously "fill in" after damage? This study doesn't yet answer that.
Summary
Understanding breakdown is how we better understand intelligence. 🎞️
This paper tells us: AI is not an inscrutable black box. Its language abilities have clear divisions of labor — logic is logic, vocabulary is vocabulary, and the foundation is the foundation.
Through this "self-injury" style of research, we are inching toward the ultimate question: how does a string of 0s and 1s weave, in the folds of a digital brain, something indistinguishable from human emotion and thought?
The deepest insights often come from staring at what is broken. 🧠✨
*(Note: this is an editorial translation/summary of the original zhichai.net forum post; the paper title and date are reproduced as stated in the source.)*