Language models resemble more than just language cortex, show neuroscientists

In a paper presented in November 2025 at the Empirical Methods in Natural Language Processing (EMNLP) conference, researchers at the Swiss Federal Institute of Technology (EPFL), the Massachusetts Institute of Technology (MIT), and Georgia Tech revisited earlier findings that showed that language models, the engines of commercial AI chatbots, show strong signal correlations with the human language network, the region of the brain responsible for processing language. 

In their new results, they found that signal correlations between model and brain region change significantly over the course of the 'training' process, where models are taught to autocomplete as many as trillions of elided words (or sub-words, known as tokens) from text passages. 

The correlations between the signals in the model and the signals in the language network reach their highest levels relatively early on in training. While further training continues to improve the functional performance of the models, it does not increase the correlations with the language network. 

The results lend clarity to the surprising picture that has been emerging from the last decade of neuroscience research: That AI programs can show strong resemblances to large-scale brain regions—performing similar functions, and doing so using highly similar signal patterns. 

Such resemblances have been exploited by neuroscientists to make much better models of cortical regions. Perhaps more importantly, the links between AI and cortex provide an interpretation of commercial AI technology as being profoundly brain-like, validating both its capabilities as well as the risks it might pose for society as the first ever synthetic braintech. 

"It is something we, as a community, need to think about a lot more," said Badr AlKhamissi, doctoral student in neuroscience at EPFL and first author of the preprint, in an interview with Foom. "These models are getting better and better every day. And their similarity to the brain [or brain regions] is also getting better—probably. We're not 100% sure about it."

From vision to language

Surprising correspondences between brain regions and AI programs were first discovered in 2014, by researchers studying a very different brain region: The visual cortex. As its name suggests, the visual cortex is known to be responsible for performing visual cognitive functions, like recognizing objects.

Prior to 2014, neuroscientists had already learned that it was patterns of bioelectrical signals within the visual cortex that enabled such behaviors. However, while they could measure the signal patterns responsible for recognizing objects, and determine the subregions where the most important signals happened, they could not predict them. The signals therefore remained mysterious. 

As it turned out, there existed a relatively simple method to make progress. This was to create a computer program, also composed of many independent signal processing units, and try to optimize it to perform the same flagship tasks of the visual cortex, like recognizing objects. 

Around 2012, such optimizations were first proven to be possible for the type of signal processing program called an artificial neural network (ANN), which intriguingly, had also been invented in the 1940s as an abstract model of brain tissues. 

Surprisingly, neuroscientists showed that these new, highly optimized variants of ANNs showed signal patterns highly correlated with those measured from the visual cortex. These programs therefore provided a putative explanation of how and why the signals happened. Some researchers also began to think of these AI programs as models of the visual cortex—undoubtedly incomplete in many ways, but still far better than any previously invented.

The question for language neuroscientists was whether similar methods could be applied to understand how the brain processed language. In particular, they wondered whether they could explain data taken from magnetic resonance imaging (MRI) experiments. One set of MRI data, examined in the recent results, was published by the researcher Francisco Pereira and colleagues in 2018. It is helpful to have a concrete understanding of this data to understand the new results. 

In one experiment from Pereira et al's study, human participants were asked to read 384 sentences from 96 different text passages. For example, "Beekeeping encourages the conservation of local habitats." Each sentence was presented on a screen to each participant one at a time, for four seconds, followed by four seconds of 'fixation,' where the activity in their brain was measured—in particular, the spatial distribution of activity across their entire brain volume. This provided a snapshot of the way the brain was spatially executing cognitive behaviors associated with reading, like comprehending language.

In a notable study from 2020, co-authored by Martin Schrimpf of EPFL, who was also a co-author of the recent EMNLP results, researchers sought to predict the patterns measured from the Pereira dataset. In particular, they sought to predict the patterns that emerged in the subregion of the brain known to be specialized for language. This region is called the language 'network,' because unlike the visual cortex, it is composed of areas that are spatially discontiguous.

In an echo of earlier results on vision, they found that for certain language models, when they processed the same sentences as the participants, their internal patterns of signals could be highly correlated with the spatial patterns of activity measured in the human language networks. These correlations were strong and significant. After creating linear regression maps between the signals in a model and in a participant, you could predict the new patterns in a participant, from a model, about just as well as you could predict the patterns from another human participant. 

The conclusion was there appeared to be deep commonalities in the ways that language models and the language network processed language. 

However, progress elsewhere in AI was making the reality appear more complicated. Commercial language models were increasingly performing tasks like mathematical reasoning, which neuroscientists think of as being highly distinctive—performed by an entirely separate brain region, called the 'multiple demand' network, recruited for solving multiple different kinds of problems. This made neuroscientists wonder how closely language models should map to merely the language network, or whether they might map to other regions, also.

More than cortex

This observation motivated the authors of the recent EMNLP results to take a closer look at the correlations between a language model and the language brain region. In particular, they sought to track how these correlations changed over the course of a model's training process, when a model is known to go from bad at language to being good at many tasks. 

They found that as a model is trained, its signal correlations with the language network peak after a relatively small amount of training tokens (two to eight billion). From there, while the model continues to grow in capabilities, allowing it to pass more language processing benchmarks, its correlations with the language network do not significantly increase (or decrease). 

Presumably, during this latter stage of training, the model might develop more correlations with other brain regions. According to AlKhamissi, the data on those other regions that would allow for making such comparisons is actively being collected.

It is tempting to use such findings to help interpret commercial AI technology, which is often thought of as being a black-box that is hard to reason about. If you were to choose to do so, they would suggest that a modern, industrial language model is less analogous to a language network—and more analogous to a set of multiple synthetic cortical regions; like a language network plus a multiple demand network—and possibly others. In other words: Less than a brain, more than a cortex. 

The field of neuroscience tends to be conservative in making such comparisons. There are many known limitations with seeing AI programs as models of brain regions, even those that have high signal correlations. For example, such models lack any direct implementations of biochemical signalling, which is known to be important for the functioning of nervous systems.

However, if such comparisons are valid, then they would suggest, somewhat dramatically, that we are increasingly surrounded by a synthetic braintech. A technology not just as capable as the human brain, in some ways, but actually made up of similar components.

Author's disclosure: No AI was used in writing or editing this post.