For decades, the "black box" of human language processing was studied through indirect means: reading times, eye-tracking, or coarse-grained neuroimaging. Simultaneously, Artificial Intelligence developed its own "black boxes"—Large Language Models (LLMs) that can write poetry and code but whose internal logic remains opaque.
A rapidly growing research direction seeks to align these two systems. By treating LLMs as "model systems" for linguistics Mahowald and Futrell, researchers are using the internal states of AI to predict brain activity, and using brain activity to "reverse-engineer" more human-like AI.
The Geometry of Thought
The core mechanism of this research is brain alignment: the mathematical mapping between the high-dimensional vector space of an LLM and the neural activity patterns (voxels in fMRI or sensors in EEG/MEG) of a human reading or listening.
Recent evidence suggests that as LLMs scale up in size and training data, their internal "geometry" for representing words naturally converges toward the patterns found in the human brain Li. This convergence isn't just a byproduct of performance; Jin and Xiong found that techniques like Reinforcement Learning from Human Feedback (RLHF), which are designed to make AI more helpful to humans, also significantly boost the model's "brain-likeness" Jin and Xiong.
Formal vs Functional Competence
However, alignment is not a linear path toward a perfect digital brain. Researchers have identified a critical distinction between formal linguistic competence (the rules of grammar) and functional competence (reasoning and world knowledge).
While LLMs are excellent at mirroring the brain's "language network"—the regions responsible for syntax and structure—they often diverge when it comes to social intelligence or physical commonsense AlKhamissi. In fact, An et al. discovered a startling paradox: as models like Llama 3.1 become more powerful at predicting the next word, they may actually become less like the human brain in how they represent hierarchical grammar An et al.. This suggests that AI may be finding "non-biological" shortcuts to achieve high performance.
The Temporal Gap
If the "what" of language is converging, the "how" remains distinct. Humans process language as a continuous, iterative stream of recalibration. In contrast, LLMs process information in "discrete bursts" at the end of their transformer layers Zhou et al..
To bridge this, researchers are using tools like the Hierarchical Frequency Tagging Probe (HFTP) to track the rhythmic signatures of grammar processing in both systems An et al.. Others are fine-tuning models to match human "surprisal"—the degree of unexpectedness in a sentence—to better predict human reading times Kiegeland et al.. Interestingly, making a model a better predictor of human behavior often makes it worse at standard NLP tasks, revealing a fundamental trade-off between human-like processing and raw engineering efficiency.
Beyond the English-Centric Bias
A major hurdle in this field is the "English-centric" nature of both AI training and neuroimaging data. Because different languages may activate distinct neural pathways—for instance, the brain adapts its parsing strategy based on whether a language is "right-branching" like English or "left-branching" like Chinese Wang et al.—researchers are pushing for more diverse benchmarks.
New datasets like MulCogBench Wang et al. and ChineseEEG Mou are providing the necessary "wetware" data to test if brain-AI alignment is a universal phenomenon or merely an artifact of English grammar. These studies are also exploring how bilingualism reshapes the brain's architecture, suggesting that "human-like" is not a single target but a spectrum of cognitive profiles Liu et al..
Why This Matters
The goal of this alignment isn't just to build better chatbots. By "brain-tuning" models to jointly predict neural activity from multiple people, researchers like Moussa and Toneva are creating AI that can generalize across individuals, potentially leading to more robust Brain-Computer Interfaces (BCIs) Moussa and Toneva. Ultimately, this research treats the LLM as a mirror: by understanding where the reflection breaks, we learn what makes human cognition unique.