Summary: A high-resolution neuroimaging study shows that both the human brain and large language models (LLMs) use deeply parallel prediction strategies to process language. Using continuous audiobook listening with simultaneous millisecond-scale EEG and MEG recordings, researchers found that the brain pre-activates representations of upcoming words: predictable words produce reduced neural activity while unexpected words trigger pronounced neural responses. The results reveal a close alignment between human predictive processing and the probabilistic predictions generated by modern LLMs, with implications for diagnostics, brain-computer interfaces, and therapies.
Although biological brains and digital models operate on very different physical media, the data indicate a convergence in the internal organization of language representations—both systems appear to construct highly parallel structural maps that support prediction and comprehension.
Key Findings
- Measurable Prediction: High-density EEG and MEG demonstrated that the human brain becomes active milliseconds before an expected word begins.
- Inverse Response Strength: Neural amplitudes scale inversely with word predictability: highly expected words elicit smaller responses, while surprising words cause stronger neural spikes.
- Parallel Structural Maps: Despite different substrates, both biological and artificial systems appear to organize language via parallel internal structures that facilitate prediction.
- Naturalistic Stimuli: Continuous audiobook listening allowed researchers to observe prediction and processing in realistic, flowing language rather than in isolated sentences.
- Clinical and Technological Potential: Mapping predictive processes at high resolution suggests new paths for diagnosing language-processing disorders, improving speech therapy, and designing higher-fidelity brain-computer interfaces.
Source: FAU
Are humans born with an innate grammatical framework, or does language emerge from use and experience?
This long-standing question in linguistics gains new perspective from research comparing human brain activity with predictions made by powerful AI language models. These LLMs operate by estimating the probability of the next word, a mechanism that makes them suitable tools for probing predictive language processing in the brain.
“We combined continuous audiobook listening with simultaneous EEG and MEG measurements and directly compared participants’ brain activity to prediction probabilities from large language models at millisecond resolution,” explains Dr. Patrick Krauss.
Are the brain’s predictions measurable?
Yes. The recordings show anticipatory neural activity appearing before the acoustic onset of a word. When a word is highly predictable in its context, the later recognition response is reduced; conversely, unexpected words produce stronger neural reactions. These patterns mirror the probabilistic predictions LLMs assign to words, indicating that the brain actively predicts language in a way that can be quantified and compared to computational models.
Artificial language models are implemented as numerical neural networks inspired by the brain’s architecture, but they operate with mathematical operations on digital hardware rather than electrochemical signaling. Still, the study found surprising parallels: not only do brain and models produce similar predictive outcomes, they also seem to organize internal linguistic information in comparable ways.
Do our brains and AI work on similar principles?
The study supports core assumptions from cognitive neuroscience about predictive coding and simultaneously helps explain why LLMs perform well across many language tasks. While similar outputs do not prove identical mechanisms, the results suggest both systems follow related information-processing principles—parallel computations that integrate prior context with incoming signals to minimize surprise.
“The similarity in results does not mean the two systems are the same,” notes Achim Schilling. “But it does raise compelling questions about why such different substrates converge on comparable organization and where the limits of that convergence lie,” adds Patrick Krauss.
What comes next?
The research team plans to test whether the discovered principles are robust across different listeners, languages, and tasks, and whether they can be translated into practical applications. A clearer mapping of how the brain and LLMs represent and predict language could inform new diagnostic tools, personalized rehabilitation strategies, more effective brain-computer interfaces, and approaches to develop more interpretable AI.
Key Questions Answered:
A: They combined electroencephalography (EEG) and magnetoencephalography (MEG) to capture electrical and magnetic brain activity at millisecond resolution while participants listened to a continuous audiobook. This setup preserved natural language flow and revealed neural shifts that occurred before word onsets, demonstrating ongoing predictive processing.
A: When a word violates prediction, the brain generates a larger neural response compared with expected words. This amplified activity appears to be an error-correction signal: the system quickly updates its internal model to incorporate the surprising input, which requires additional processing resources.
A: No. Similar computational outcomes do not imply similar subjective experience or consciousness. Human brains operate via electrochemical signaling, neurotransmitters, and plastic biological networks, while LLMs compute probabilities and vector transformations on digital hardware. The study indicates shared mathematical principles, not shared consciousness or identical mechanisms.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full by the editorial team.
- Additional context was provided by staff.
About this research
Author: Doha El Ezzi
Source: FAU
Contact: Doha El Ezzi – FAU
Image: Image credit: Neuroscience News
Original Research: Open access. “The predictive brain: Neural correlates of word expectancy align with large language model prediction probabilities” by Kölbl N, Tziridis K, Maier A, Kinfe T, Chavarriaga R, Schilling A, Krauss P. NeuroImage. DOI: 10.1016/j.neuroimage.2026.121966
Abstract
The predictive brain: Neural correlates of word expectancy align with large language model prediction probabilities
Predictive coding theory proposes that the brain continuously anticipates upcoming words to optimize language processing, but the neural underpinnings remain unclear for natural speech. In this study, EEG and MEG data were recorded simultaneously from 29 participants listening to an audiobook while predictability scores for nouns were computed using three language models (one BERT and two multilingual LLaMA variants).
Results indicate that higher predictability correlates with reduced neural responses during word recognition, observable as smaller N400 amplitudes, and with increased anticipatory activity prior to word onset. EEG pointed to pre-activation in left fronto-temporal regions, while MEG suggested greater sensorimotor engagement for low-predictability words, hinting at a possible motor-related contribution to linguistic anticipation.
These findings show dynamic integration of top-down predictions and bottom-up sensory input during naturalistic language comprehension. To the authors’ knowledge, this is the first study demonstrating these effects in continuous speech, linking computational language model predictions with high-resolution neurophysiological measures. The results advance our understanding of predictive processing in language and may inspire neuroscience-informed AI development.