Researchers Use Lesions to Study AI Language Errors

University of South Carolina scientists intentionally damaged large language models to observe errors resembling those seen in aphasia patients.

Updated on Nov. 20, 2026 in Artificial Intelligence

Isometric editorial illustration of a copper cable bundle with one section cut open and its inner wires splayed across a slate surface.
University of South Carolina researchers deliberately damaged large language models and found their errors resembled the language deficits seen in aphasia patients. AI Illustration. Upload story photo >

Live Poll

Should the government fund AI research aimed at helping stroke survivors regain speech?

Researchers at the University of South Carolina introduced deliberate errors, called lesions, into large language models and found the models produced language processing mistakes similar to those of people with aphasia. The method, called lesion-based interpretability, establishes cause and effect rather than the correlational comparisons used in earlier studies.

Why it matters

Because language models are trained on human language and aphasia is an impairment of human language, the approach offers a controlled way to study how language breaks down. The researchers say the framework could also be applied to neurological conditions beyond aphasia.

The study perturbed layers within language models and administered clinical tests used on aphasia patients to compare failure patterns. The extraction does not report specific layer counts, test scores or sample sizes, and it is unresolved how far the model structures mirror the human brain.

The players

University of South Carolina

A public research university in Columbia, South Carolina, whose researchers conducted the study on lesion-based interpretability of language models.

The details

Researchers intentionally introduced lesions into the models, then administered clinical tests used on aphasia patients to compare the resulting errors. The paper reports correspondence between patterns of failure, not between the actual structures of the models and the brain.

The Tech Race

The work extends the lesion method neuroscientists have long used to map human language, applying it to artificial systems instead of injured brains. Its authors frame the approach as a route to smaller, more efficient language models, though that projection is not yet demonstrated.

The study is at the research stage and does not yet change any product people use. Its projected uses, including smaller models and digital twins of aphasia patients for simulated clinical trials, remain future possibilities.

The takeaway

Studying AI systems through deliberate damage may give researchers a clearer view of how language breaks down, in machines and in people. The approach is promising, but its clinical value is still to be tested.

Further reading

For more on the field this study draws on, explore our Artificial Intelligence coverage.

Source note: This article includes information reported by Labcompare.

Live Poll

Should the government fund AI research aimed at helping stroke survivors regain speech?