Two recent studies suggest that the large language models behind AI chatbots resemble people in more ways than their fluent output. One, published in Nature Communications, finds that the layered structure of these models lines up with the order in which the human brain makes sense of spoken language. The other, posted to the preprint server arXiv, finds that the same models reproduce the “us versus them” group biases that psychologists have long documented in humans.
Brain activity tracks the layers of a language model
The Nature Communications study was led by Ariel Goldstein of the Hebrew University of Jerusalem, with Mariano Schain of Google Research and collaborators at Princeton University. The team recorded brain signals from participants while they listened to a roughly 30-minute podcast, then compared those signals with the internal representations produced by a large language model processing the same words.
According to the paper, the brain’s response unfolds in stages. The earliest activity after each word reflects basic features of the sound and the word itself, which corresponds to what the early layers of a language model compute. Later activity matches the deeper layers of the model, where context is integrated and meaning takes shape. The authors report that this alignment was especially clear in Broca’s area, a region of the frontal lobe central to speech and language.
The researchers describe the result as evidence that the model’s layer hierarchy aligns with the temporal dynamics of language comprehension in the brain. In their view, this challenges older accounts that treat human language understanding mainly as the application of symbolic rules, and it supports the idea that the brain, like the model, builds meaning gradually through a cascade of increasingly abstract representations.
The finding does not mean the brain is a language model, and the authors do not claim that. The comparison is statistical: patterns in neural recordings predicted by patterns inside the model. But it adds to a growing body of work using modern AI systems as tools for studying how biological language processing works.
Models inherit human group bias
The second study, titled “Us-vs-Them bias in Large Language Models” and listed on arXiv in December 2025, approaches the human comparison from a different angle. Drawing on Social Identity Theory, the framework psychologists use to explain favoritism toward one’s own group and hostility toward outsiders, the authors tested several large language models under default conditions and after assigning them a persona.
The paper reports that the models displayed in-group favoritism and out-group negativity in a pattern similar to what is observed in people. When models were given a political identity, their responses shifted in ways that echoed partisan behavior in human studies. The authors interpret this as a sign that models absorb not only facts from their training data but also social attitudes and ways of framing other groups.
The study also tests a mitigation approach and reports that it substantially reduced the measured bias, suggesting that the behavior, while present by default, is not fixed.
Why the comparison matters
Taken together, the two papers point in the same direction: systems trained to predict human text end up reflecting human cognition, both its structure and its flaws. The first study shows the resemblance at the level of how language is processed; the second shows it at the level of social judgment.
As language models are built into tools for hiring, customer service, education and public information, the authors of the bias study argue that understanding these inherited tendencies is a practical concern rather than a theoretical one. A model that mirrors human in-group bias could reinforce it at scale, and the mitigation results suggest developers have room to counteract it if they measure for it.
