Researchers deliberately disrupted large language models, gave them clinical language tests and found errors resembling those of patients with aphasia, according to a Cell Press announcement published November 19 on Tech Xplore. The study, published in Patterns, uses methods from clinical neurology to investigate AI's language failures. Getting the model to fail was part of the plan. Source
Corresponding author Julius Fridriksson of the University of South Carolina says the findings concern similar patterns of failure, not corresponding physical structures. Matching mistakes don't establish matching machinery. Source
Fridriksson's lab studies aphasia—difficulties with language and speech—including in people who have experienced strokes. The team calls its approach to investigating how language models work “lesion-based interpretability.” Source
“So, the question was whether we could perturb layers in the language model, give it the same clinical tests we give our patients, and see whether the output looked like aphasia,” Fridriksson says in the announcement. Source
The team's proposed applications include identifying redundant components to develop smaller, more energy-efficient models and creating “digital twins” of patients with aphasia to simulate clinical trials. The announcement reports no demonstrated energy savings or patient benefits. Source
The Cell Press account does not identify the models, patient sample size, specific clinical tests or numerical strength of the similarities. Without those details, the announcement alone doesn't establish how robust or broadly applicable the match is. Source
Comments