Plenary Panel Session:

Human Language and Thought in the Era of Large Language Models

Friday, May 23, 6:00 PM – 7:30 PM 

This panel brings together leading early career voices whose work  
engages with the relationship between human language and LLMs. How do these advances in language technologies influence how we think about and study the psychology and neurobiology of human language processing?  

The capacity to rapidly communicate and think with language is a remarkable human ability that has been argued to separate humans from other forms of intelligence, biological or artificial. However, recent advances in artificial intelligence, specifically large language models (LLMs), appear to be approaching human-like performance. This symposium brings together leading early career voices whose work engages with the relationship between human language and LLMs. How do these advances in language technologies influence how we think about and study the psychology and neurobiology of human language processing?  

Chair: APS President Randi C. Martin, Rice University 

Mapping and Decoding Language Representations from Human Cortex  

Is it possible to read the content of human thought out using recordings of brain activity? We use non-invasive functional MRI and machine learning methods based on large language models to investigate the relationship between brain activity and the content of thought. These methods reveal complex spatial and temporal patterns of activity that relate to specific semantic categories. We show that this information can be read out as language, even when the stimulus evoking it is from another modality. We also use self-supervised speech models to show that fine temporal information can be deduced even from slow non-invasive recordings. These results point to a future of neuroscience that strongly integrates modern neural network models.

Speaker: Alexander Huth, The University of Texas at Austin 

Dr. Alexander Huth, Assistant Professor of Neuroscience and Computer Science, University of Texas, Austin, is an expert in using quantitative, computational methods to understand how the human brain represents meaning. His work focuses on how to build encoding models, like neural network language models, that predict human brain responses while people listen to podcasts or watch movies, with the goal of understanding how language is represented across the brain. He is exploring potential applications for decoding intended messages from the brain signals of those with disrupted language following brain damage. 

Neural Algorithms of Human Language 

For the first time in history, there exist systems other than the human brain that can process speech and language, extract meaningful symbolic structure, and produce complex and appropriate responses. Laura Gwilliams will present studies from her lab that use these large speech and language models to generate algorithmic hypotheses of the biological implementation of language understanding. The work uses neural timeseries data across different spatial scales: From population ensembles using MEG and intracranial EEG, to the encoding of speech properties in individual neurons across the cortical depth using Neuropixels probes in humans. The results provide insight into what representations and operations serve to bridge between sound and meaning in biological and artificial systems, including how information at different timescales is nested, in time and in space, to allow information exchange across hierarchical structures. Together, the findings represent a new era of scientific inquiry to understand system-level implementations of human language. 

Speaker: Laura Gwilliams, Stanford University 

Dr. Laura Gwilliams, Assistant Professor of Psychology, Stanford University, is an expert in understanding the neural mechanisms that underlie our ability to comprehend speech. Her work focuses on understanding the representations the brain derives from auditory input and the computations that are applied to those representations to allow us to understand meaning, to inform both our understanding of the human brain and our ability to build intelligent machines. 

Dissociating Language and Thought in Humans and in Machines 

Today’s large language models (LLMs) routinely generate coherent, grammatical, and seemingly meaningful paragraphs of text. This achievement has led to speculation that LLMs have become “thinking machines,” capable of performing tasks that require reasoning and/or world knowledge. Anna Ivanova will discuss how easy it is to conflate language and thinking, both in humans and in machines. To address this conflation, she will introduce a distinction between formal competence—knowledge of linguistic rules and patterns—and functional competence—understanding and using language in the world. This distinction is grounded in human neuroscience, which shows that formal and functional competence recruit different brain mechanisms. Ivanova will then discuss how researchers can leverage behavioral and neuroscience approaches from the study of human intelligence to carefully examine—and dissociate—distinct capabilities in artificial intelligence (AI) systems, and, in turn, how advances in AI can contribute to our understanding of language and cognition in humans.  

Speaker: Anna Ivanova, Georgia Institute of Technology 

Dr. Anna Ivanova, Assistant Professor of Psychology, Georgia Tech, is an expert in understanding the relationship between language, intelligence, and human thought. Her work focuses on the interaction between the neural circuits responsible for language and thought, the relationship between large language models and human cognition, and the role of inner speech on thought. 

Moderator: L. Robert Slevc, University of Maryland, College Park 

L. Robert Slevc, Associate Professor, Psychology, University of Maryland. His expertise is in the processing of language and music and the extent to which the same brain regions support both.