Biological prediction machine: Our brain works similarly to artificial intelligence when it comes to understanding language – it uses probabilities to predict the next word, as an experiment demonstrates. Brain activity therefore reflects how semantically likely a word heard is. This suggests that AI models and the human brain follow similar principles when processing language. This provides new approaches for both neuroscience and AI research.
When it comes to how our brain works, there are still many unanswered questions. But one aspect is becoming increasingly clear: “The human brain is a prediction machine – it continuously anticipates sensory stimuli, words, events and, in general, future conditions,” explain Nikola Kölbl from the University of Erlangen-Nuremberg and her colleagues. This ability to anticipate possible developments enables us to react quickly and ensured the survival of our early ancestors.
A look into the brain while listening to audio books
But it is still unclear whether our brain also uses this predictive coding for language processing. “Despite decades of research, the neural mechanisms behind the extraction and representation of word meanings are still not fully understood,” explain Kölbl and her team. Advances in artificial intelligence have brought this topic into focus again because it raises the question of how similar large language models such as GPT, Gemini or Claude and our brain are when it comes to language processing.
To clarify this, Kölbl and her colleagues examined how the human brain processes heard language – and whether it works with predictions and probabilities. To do this, they analyzed the brain reactions of 29 test subjects while they listened to an audio book. “In our study, we combined the natural, continuous speech of an audio book with simultaneous electroencephalographic and magnetoencephalographic measurements,” explains senior author Patrick Krauss from the University of Erlangen-Nuremberg.
The highlight: The team compared the test subjects’ brain activity directly with the prediction probabilities of several large language models such as Llama, GPT-4o and BERT for the same texts.
Probability determines brain activity
And indeed: The evaluations revealed surprisingly clear parallels between AI and the human brain. Our thinking organ also tries to anticipate the next words when listening, as the brain activity of the test subjects revealed: the more likely a word is in the respective context, the weaker the neural reaction will be when processing the word heard. In contrast, unexpected words trigger stronger neural responses.
“This is consistent with the idea that the brain has to work harder when it encounters unexpected words,” explains the team. At the same time, the data showed that brain activity in the language-processing brain regions increases shortly before the next word begins. This suggests that the brain makes semantic predictions based on the words it has already heard – it determines which word is most likely to follow next. “We were able to demonstrate that the brain actively predicts language,” says Krauss.
Parallels between the brain and AI
This means: When it comes to language processing, our brain and common AI models don’t work all that differently. Both use learned probabilities to predict word order. “We were particularly surprised that not only similar predictions emerge between the brain and language models. There is increasing evidence that both systems organize language internally in a comparable way,” says Krauss.
The study results thus support central assumptions of cognitive neuroscience and at the same time provide an explanation for why AI language models are so powerful in many applications.
How far do the similarities extend?
However: “The fact that the brain and language models come to similar results does not automatically mean that they function in the same way,” says Kölbl’s colleague Achim Schilling. Artificial intelligence is based on mathematical information processing units whose architecture is modeled on the human brain, but which work with algorithms and numerical values instead of cells, biochemical and electrical signals.
“The exciting question is why two systems that are so different still converge on such identical forms of linguistic organization – and also where the limits of this convergence lie,” says Krauss. “If we better understand how the brain and language models represent and predict language, in the long term this could lead to new approaches for diagnostics, personalized therapies, brain-computer interfaces or better explainable AI.”
Source: Nikola Kölbl (Friedrich Alexander University Erlangen-Nuremberg) et al., NeuroImage, 2026; doi: 10.1016/j.neuroimage.2026.121966