A person who can no longer form a sentence after a stroke may still solve a difficult puzzle, spot a hidden pattern or make a sound logical judgment. Researchers at MIT's McGovern Institute for Brain Research have now shown that severe language impairment does not prevent logical reasoning — and that the brain's language system is not even recruited when people reason logically.
The study, published in the Proceedings of the National Academy of Sciences and led by Evelina Fedorenko, examined two people whose strokes had severely damaged the brain's language regions. Both had great difficulty understanding and producing language. Yet they performed as well as control participants on language-free logic puzzles — even as the tasks became more demanding: they uncovered hidden rules in lists of numbers, applied them to new examples and completed matrices of geometric patterns. They communicated the rules they inferred through gestures and drawings.
'It really upends a theory that says that symbolic rule induction is not possible without linguistic capacities,' says lead author Hope Kean.
Functional brain scans of healthy adults provided the second line of evidence: neither inductive nor deductive reasoning activated the brain's language regions. The 'multiple demand network,' a distributed system supporting complex problem-solving, did respond during inductive reasoning but, unexpectedly, not during deductive reasoning — a difference Kean is continuing to investigate.
The work carries practical weight for how aphasia is perceived: language loss is often mistaken for lost intelligence, even though people with aphasia may still manage finances, play chess or sudoku and make complex decisions. 'We should continue to educate the public that linguistic difficulties — in aphasia, but also in those with developmental language conditions, such as stuttering, or those who do not speak English natively — are not indicative of how smart or capable someone is,' Fedorenko says.
The findings also inform AI research. Large language models learn from text and generate text, and they can imitate certain forms of human reasoning. If language and abstract logic operate separately in the human brain, that helps researchers understand what current models are actually doing — and how future systems might be designed.
