ChatGPT: A Tool for Understanding Human Brain Language Processing

Can researchers learn more about how the human brain processes language by using ChatGPT?

Surprisingly well, large language models can replicate human speech and writing. They are now acting as an electronic laboratory rat for language, providing light on the riddles surrounding one of the most significant capacities of humanity.

Toward the end of 2018, MIT neuroscientist Evelina Fedorenko started hearing about a new wave of “large language models” (or LLMs) emerging from Silicon Valley's AI labs. When she saw outputs from these models—also known as “generative AI” for their ability to generate and comprehend prose—she was astounded. “It was just too good,” Fedorenko said, as the machine-generated passages sounded remarkably coherent and human-like, far beyond what she had imagined possible in her lifetime.

LLMs have provided researchers like Fedorenko with new tools to explore language processing in the brain, challenging long-standing understandings of how we process language in discrete areas like Broca’s region. As more advanced models appeared, they proved to be more than mere sophisticated tricks; their linguistic capabilities were undeniable.

By 2022, language models like ChatGPT had sparked a global sensation, demonstrating capabilities closely mirroring those of the brain's language regions. This unexpected alignment has offered neuroscientists unprecedented insights into how our brains extract meaning and structure from language, an area previously impenetrable due to the uniqueness of human language processing.

Language regions in the brain, primarily located in the left hemisphere, are specialized for understanding and producing language. Functional MRI (fMRI) studies have shown that these areas activate consistently for language tasks but not for abstract forms of cognition such as logic or music. These regions form part of a network that collaborates with other brain systems to process extended discourse, yet traditional methods like fMRI have limitations in revealing the precise mechanisms at play.

LLMs, on the other hand, provide a novel "electronic lab rat" for language research. They have demonstrated that prediction plays a crucial role in language processing, both in AI and the human brain. The better an AI model predicts the next word, the more its activity aligns with brain data, supporting the predictive brain hypothesis. This theory suggests that the brain constantly anticipates incoming information and refines its expectations based on actual input.

Recent studies have used AI models to uncover how the brain's language regions react to different types of sentences, revealing that simple, grammatically correct sentences elicit low neural activity, while more complex or unusual sentences require greater cognitive effort. These findings highlight the brain's efficiency in processing straightforward language while dedicating more resources to interpreting ambiguous or complex inputs.

Researchers like Fedorenko and her team are leveraging these AI models to further investigate the intricacies of language processing. By manipulating the models and comparing their responses to human brain activity, they hope to map out how the brain integrates words into sentences and interacts with other cognitive functions. This approach promises to enhance our understanding of the brain's language network and could lead to breakthroughs in how we model cognitive skills in AI.

In summary, the alignment between AI language models and the brain's language regions offers a promising avenue for exploring the neural mechanisms of language. As researchers continue to use these models to probe the brain's functions, they may uncover new ways of understanding how we process and produce language, with potential applications in both neuroscience and artificial intelligence.