Researchers Replicate Brain Lesions in AI Models to Study Language Mechanisms
In a study published in Patterns, University of South Carolina researchers introduced artificial lesions into large language models, and the damaged models made errors resembling those of human aphasia patients.
Newsnibs Staff
- 2026-11-22
- Updated 7:00 AM UTC
Three separate studies on language and the brain have come out from different teams. They are not reports of a single project. One damages AI language models the way a stroke damages a brain. One challenges the textbook account of where speech is produced. One uses machine learning to flag language difficulties in brain scans.
The first comes from the University of South Carolina, where a team led by Julius Fridriksson published in the journal Patterns. Fridriksson's lab specializes in aphasia, a language disorder often caused by stroke. The researchers introduced artificial "lesions," or errors, into specific layers of large language models. They call the approach lesion-based interpretability. It is modeled on more than 150 years of using brain-lesion analysis to learn which areas support which language functions. According to co-author Roger Newman-Norlund, the damaged models made language and cognitive errors very similar to those of human aphasia patients. The authors state that this does not imply the models have structures equivalent to the brain's.
Fridriksson says it is the first study to apply clinical lesion methods to AI, and that it aims to establish cause and effect rather than correlation alone. He notes that language models are already used in high-stakes settings even though their inner workings are opaque. The team hopes the method will help explain why the models fail. It could also identify redundant components, which would allow smaller and more energy-efficient models. The same approach could be used to build digital twins of aphasia patients that simulate clinical trials and speed discoveries into care.
The second study was presented at the annual meeting of the Society for Neuroscience by a University of California, San Francisco team led by neurosurgeon Edward Chang. The team used direct recordings from epilepsy patients to map language activity. Their findings challenge a consensus of roughly 150 years that Broca's area, in the frontal lobe, is the main region for producing speech. They point instead to the lpSTG, a region within Wernicke's area in the temporal lobe. It generates "auditory images" of planned speech and sends those signals directly to the motor centers that control the vocal tract. On this account, the temporal lobe is the main driver of language production.
The work bears on brain-computer interfaces for people with aphasia. Decoding signals from upstream planning areas such as the lpSTG could allow communication when the motor centers cannot be used. John Jay Ngai, director of the NIH's BRAIN Initiative, called the study an "amazing achievement."
The study leaves the role of Broca's area unresolved. Evelina Fedorenko, a cognitive neuroscientist at MIT, asked why the frontal-lobe parts of the language network exist at all if the temporal lobe is the main center of action.
The third study was published on Friday by researchers at Beijing Technology and Business University and Beijing Wuzi University. They combined explainable AI with multimodal EEG and fMRI biomarkers to identify language difficulties. The data included 129-channel EEG recorded at 500 Hz and fMRI from 1.5T and 3T Siemens scanners. Funding came from the National Natural Science Foundation of China and the Beijing Scholars Program. A strictly nested cross-validation procedure was used to prevent label leakage. In analyses adjusted for covariates, 37 of the 38 multimodal features stayed significant after false-discovery-rate correction. The exception was the standard deviation of Global Field Power for Microstate E. The stability of the feature rankings was assessed across 10 runs of 5-fold cross-validation.
Photo by Shawn Day on Unsplash