Quanta recently had a conversation with Fedorenko regarding the language network in our brains, likening it to the digestive system. They discussed her insights on how the language decoder operates and whether she thinks people possess large language models (LLMs) in their minds. The dialogue has been simplified for better understanding.
What is the language network?
In adult brains, there exists a primary set of areas that work together as an intricate system for understanding linguistic structure. These areas keep track of the connections between words and meanings, along with the rules that dictate how to combine words. When you learn a new language, you’re really picking up these connections and rules, which in turn enables us to use this “code” in remarkably flexible ways. Essentially, you can switch between thoughts and word sequences in any language you know.
That sounds very abstract. But you call the language network a “natural kind” — does that mean it’s something physical you can point to, like the digestive system?
Exactly. The systems identified in the brain, such as the language network and certain visual areas, resemble organs in some senses. For instance, the fusiform face area is very much a defined unit. In the case of the language network, there are generally three regions in the frontal cortex for most people, all situated on the left side. Additionally, there are some areas along the middle temporal gyrus, spanning the temporal lobe. Those are the fundamental regions.
You can observe their unity in various ways. For example, when individuals are scanned using functional magnetic resonance imaging (fMRI), the responses to language consistently activate these regions. We’ve analyzed about 1,400 people now, allowing us to create a probabilistic map that predicts where these regions are likely located. While individual brain anatomy can vary somewhat, the general patterns tend to be quite stable. Within those broad frontal and temporal areas, there’s almost certainly some tissue responsible for linguistic computations in everyone.
How is this different from other parts of brain anatomy known to be associated with language, such as Broca’s area?
Broca’s area is quite contentious. I wouldn’t categorize it strictly as a language region; instead, it’s more of an area that coordinates motor planning for speech. In this moment, it’s helping me plan how my mouth muscles should move to articulate what I’m saying. Interestingly, I could produce a sequence of nonsensical words, and it would still activate. It processes the sounds of speech and figures out the necessary motor movements, functioning downstream from the language network.
You’ve also mentioned that language isn’t the same as thought. So if the language network isn’t generating speech and isn’t involved in thought, what is its role?
The language network acts as a bridge between lower-level perceptual and motor functions and the more abstract representations of meaning and reasoning. Essentially, language production involves taking a vague thought, along with a collection of vocabulary—not just individual words but also larger constructs and rules for their arrangement. You search through this lexicon to find a way to communicate your intended meaning through a structured sequence of words. After formulating that message, the motor system is engaged to vocalize, write, or sign it.
