“I have no words,” people often say when they feel or think something so transformative or moving that the effect seems beyond coherent expression. Love, grief, death, and awe, in particular tend to elude language’s grasp.
Poetry, which rejects many of the conventions of everyday language, may have the greatest capacity to name the unnameable. And yet, many of us think our most private thoughts in words—our inner monologues galloping an inside track through our minds all day long. So what is the role of language in thought?
This is a subject that MIT cognitive scientist and professor Evelina Fedorenko has spent decades researching. Using studies of behavior, brain imaging, intracranial recordings and stimulation, and computational modeling, she has investigated how people understand and produce language, and where language and thought overlap.
Read more: “How Does Your Brain Know a Cat Is a Cat?”
I recently interviewed Fedorenko via email about the definition of a “thought,” the role of language and culture in perception, attention and memory, and what large language models can teach us about the nature of human reasoning.
Is there such a thing as a discrete “thought” with edges, and if so how would you define it or recognize it?
This is a tricky one because it really depends on how you define “thought.” There certainly can be discrete thoughts with edges, but there also seem to sometimes be thoughts that are maybe half-formed, like partial intuitions, a glimpse of something that is hard to call discrete.
How would you recognize a discrete thought is a hard question, but one possible operational definition is that you can translate it into a symbolic form, be it a sentence or a little piece of a computer program.
What is your best hypothesis about the form of human thoughts? Are they composed of pictures, words, sounds, feelings—all of those things at once?
It’s not my hypothesis: I think this is the field’s current best guess. Much of human, and perhaps non-human, thought and reasoning can be construed as pieces of computer code, what is often referred to as “probabilistic language of thought,” or PLOT.
So, actually not pictures or words. And as a side note, I think the fact that this term “language of thought” has been in common use for many years has added to the confusion between language and thinking, but the term does not suggest that we think in language. In fact, there is now strong evidence that we don’t.
It’s also worth noting that there are many different things we think about. We think and reason about the physical world, about other people, about abstract relations—for example, imagine thinking through a way to solve a logic or a math problem. And there is now evidence that these different aspects of thought are supported by separate mechanisms in our brain.
For example, we have brain regions that help us reason about gravity, other regions that help us figure out what someone else believes or wants, and yet other regions that help us figure out how abstract elements relate to one another. Why these different capacities separate is a great question, and many of us are working to figure that out.
You authored a paper in 2024 that found that language is a tool for communication rather than thought. What was the basis for this finding, and if this is so, why then do so many of us have internal monologues?
There are two primary bodies of evidence that support the idea that we don’t use or need language to think. One comes from studies of patients with severe brain damage that affects the language areas, severe aphasia.
Some such individuals can nevertheless think just fine: They can do math, navigate in space, make social judgments, solve sudoku puzzles, and so on.
So what they lose because of their brain damage is the ability to translate their thoughts into language, and also to infer meanings from what others say or write, but they seem to preserve their ability to think.
The other line of evidence comes from brain imaging, mostly functional MRI. Work in our group and some others have shown across many different kinds of experiments that the brain areas that allow you to understand and produce language—your “language areas”— are working hard whenever we use language, but are not active when you perform various thinking tasks. So they don’t show activity when you solve a math problem or when you have to reason about someone else’s mind. Instead, other parts of the brain are hard at work.
The phenomenology of what you referred to as “internal monologue” or what is sometimes referred to as “inner speech” is super interesting and is actively studied. We know that engaging in inner speech cannot be critically important for complex thought because a non-trivial fraction of humans report no such percepts. And yet, you are right that many humans do experience it. I hope some of the ongoing work, including by Anya Ivanova at Georgia Tech, will help us understand the role inner speech may play in the minds of those who experience it often.
Some researchers have suggested that language actively influences perception, attention, and memory. For example, research on Aboriginal communities like the Kuuk Thaayorre, who speak Guugu Yimithirr, shows they use absolute cardinal directions—north, south, east, west—instead of egocentric terms like “left” and “right,” causing them to construct space and time relative to the landscape rather than their own bodies. How can we square this finding with the idea that language and thought are distinct?
I think this literature on cross-linguistic effects is fraught with confusion. Many confuse language and thinking. An even larger fraction of scientists and non-scientists alike seem to conflate language and culture. There are absolutely strong and robust effects of culture on cognition and these effects are necessarily reflected in that culture’s language, but none of the reported effects show that language affects thinking.
For example, say you live in a small farmer-forager village in Amazonia, and in your culture a certain landmark—say, a river—is really important for you to orient in space. Your thinking about space will certainly be different from someone who lives in downtown Manhattan, and the words you use to describe space will reflect the importance of the landmark.
But this is not language affecting thought, this is culture affecting your way of life and your thinking and reasoning. Language is simply a reflection of that. That’s what language does for us. It allows us to share the mental contents of our minds with each other. If I think about a banana grove being up the river, I will also talk about it in this way.
What can large language models teach us about the nature of thought?
Many things. When LLMs emerged, they nicely highlighted that language can kind of exist without the ability to think. Early models like GPT2 produced fluent language but were weak at reasoning. Of course, as models got larger and more sophisticated, they got better at many cognitive tasks.But those models are no longer just models trained on linguistic text with the objective of predicting the next word. They were also trained on massive databases of computer code—which, by the way, in the human brain is processed by different brain regions from those that process language! And often with additional objectives and with humans in the loop in some way.
So, how and why these modern models become so cognitively competent gets harder to unpack, but there is a lot of great work using mechanistic interpretability approaches to do so.
In a really exciting recent finding—this is work with Andrea de Varda, Pengrui Han, and Jacob Andreas—we see that models that are cognitively competent seem to develop separate components for language and for different aspects of thought, very much like the human brain.
And more generally, because these models are much more penetrable than the human brain (we can use all sorts of approaches that we simply can’t with current methods we have in human neuroscience), we may get some cool insights into how these models solve a math problem or reason socially, and these insights can then hopefully be tested in humans. So I think AI models will continue to inform human cognition for many years to come.
Enjoying Nautilus? Subscribe to our free newsletter.
Lead Image: HAKINMHAN / Shutterstock
