Lesion method from stroke research is brought into AI: damaged language models make patient-like errors
An aphasia research team at the University of South Carolina has transplanted neurology's classic tool into AI research: by inflicting controlled "lesions" on large language models, they report that the models make language errors resembling those of their patients. The study is published in the journal Patterns in November 2026, and according to the researchers themselves, it is the first time the lesion method has been used this way on artificial intelligence.
The news
A research group specializing in speech and language difficulties after stroke reports that the framework long used to localize language functions in the brain can also be applied to large language models (LLMs). The method has been named "lesion-based interpretability." The study is published in Patterns (2026) with DOI 10.1016/j.patter.2026.101720, and the findings were made known through a Cell Press press release that Techxplore republished on November 19, 2026.
The core reported finding is that the artificial "injuries" led to errors in language processing that resembled the errors made by the laboratory's own aphasia patients. It is nonetheless important to stress what this means and what it does not mean: the study reports a correspondence in error patterns, not that the models process language the way the brain does.
What lesion mapping is — and why it is a causal tool
In clinical neurology, the lesion method has been the workhorse for over a century. The logic is simple: when a patient suffers a circumscribed brain injury, for example after a stroke, and loses a specific function, researchers can link the lost function to the damaged brain location. Comparing many patients' injuries and functional losses yields an anatomical map of where in the brain different abilities — spoken language, word-finding, sentence formation — reside.
The method's strength is that it is causal by design: the injury is itself the cause of the functional loss. That sets it apart from most comparisons between human and machine language understanding, which have traditionally been correlational — one looks at similarities in behavior or activation patterns without intervening in the system.
This is precisely what the research group now claims to have transferred to AI. Instead of waiting for a model to fail, they induce the injury themselves, in a controlled and repeatable way, and study what happens.
How the method is applied to language models
In the new study, language models are thus given artificial "injuries," and the resulting errors in language processing are compared with the errors made by the laboratory's aphasia patients. The team says such models can help interpret what happens inside the LMs' "black box," and can be used to study model efficiency by identifying parts that are redundant and unnecessary.
The press release does not, however, state which models were tested, how many lesions were induced, which clinical tests the errors were compared against, or how well the models' errors actually matched those of the patients. That means the reader must, for now, take the method description and the similarity finding on the researchers' word — documented via the press release, not via the article itself.
The researchers' claims — with clear sourcing
Corresponding author Julius Fridriksson of the University of South Carolina frames the main point in the press release: "LLMs are being deployed in situations where failure has consequences, but we have very little ability to say why a given system fails the way it does," he says. "It turns out that methods used in clinical neurology can be applied to understand artificial intelligence. This is the first paper to use these methods in this way."
He adds: "Until now, comparisons between human and artificial intelligence language have been almost entirely correlational," and: "This approach for the first time establishes cause and effect."
Both "first" claims are the researcher's own, made in a press release with an obvious interest in favorable framing, and they have not been independently verified. They should therefore be read as the team's assessment of its own method, not as documented facts.
The built-in caveat: the same errors, not the same brain
The most important qualification comes, in fact, from the researchers themselves: the approach shows correspondence in error patterns, not in underlying structures. That a "lesioned" language model makes errors reminiscent of aphasia does not mean the model organizes language the way the brain does, or that the model's internal representations correspond to the brain's. A modern language model is trained on text statistics and has no brain anatomy to map; the similarity concerns what the systems do when they fail, not how they are built.
This differs from the historical lesion logic, where the injury's location in the brain yields direct anatomical information. In a language model, the "lesion's" location is information about the model's internal structure — what it is, for lack of a better name, not a map of anything biological.
Future applications — research speculation, not results
The press release points to two prospective uses, and both should be read as the researchers' intentions, not demonstrated results:
- Smaller, more energy-efficient LLMs. If the lesion method can identify redundant and unnecessary model parts, it could, according to the researchers, pave the way for future LLMs that are smaller and more energy-efficient. This is a hypothesis about efficiency gains, not a result the study documents.
- "Digital twins" of aphasia patients. The researchers suggest the models could also be used to create digital twins of patients with aphasia, to simulate clinical trials and translate new findings into practice more quickly. This, too, is presented as a potential in the press release.
Neither of these applications has been tested in the study as described, and both depend on the method holding up in practice — something independent replications must show.
Open questions and limits in the documentation
The public documentation of this finding comes from a single source: the Cell Press press release, republished by Techxplore. The scientific article in Patterns exists behind peer review and a DOI, but its details are not available in the source material AIMag has access to. Several concrete questions therefore remain open:
- Which language models were tested, and at what sizes?
- How many lesions were induced, and how were they chosen?
- Which language tests were used to compare the models with the patients?
- How strong is the measured similarity between the models' and the patients' errors?
In addition, the priority claims remain open: that this is "the first paper" using such methods, and that the approach "for the first time establishes cause and effect," are Fridriksson's assessments. The interpretability research field is large and fast-moving, and there are nearby techniques — including various forms of systematic deactivation of model components — that must be read and compared with this method before a priority claim can be assessed on its own.
What nonetheless makes the study worth attention is the method transfer itself: a clinical tool with a hundred years behind it is being brought into AI research, by a research group that knows exactly which error patterns real brain injury produces. Whether it yields a usable map of why language models fail — or merely a new and elegant description of the fact that they do — is something independent research must clarify.

