Study Finds Strong Similarities Between How Large Language Models and Human Brains Process Meaning and Memory

Although large language models generate text through statistical prediction, research suggests that their layered computations correspond to aspects of language processing in the human brain. The issue could prove important to efforts to reduce catastrophic forgetting and energy consumption in AI. Alphabet, Amazon, Meta and Microsoft invested nearly $400 billion in AI infrastructure in 2025, compared with just $600 million over 10 years for Astera's neuroscience program, underscoring the wide disparity in resources and the importance of the research.

On November 26, 2025, researchers from Google Research, Princeton University and other institutions published a study in Nature Communications. As nine patients listened to a 30-minute podcast, deeper representations in GPT-2 XL and Llama 2 increasingly corresponded to later responses in Broca's area. South Korea's Institute for Basic Science separately found that a Transformer simulating NMDA receptors could strengthen long-term memory.