MRI scans and tests on people with aphasia indicate that formal reasoning uses circuits distinct from the areas dedicated to words.

Speaking aloud can help organize a problem, but it doesn't mean the brain has to convert every inference into a sentence. Research led by McGovern Institute for Brain Research of the Massachusetts Institute of Technology, in United States, clearly separates the neural systems of the language from those employed in the logical reasoningThe study, published in the “Proceedings of the National Academy of Sciences", combines functional magnetic resonance imaging and behavioral testing on people with severe acquired language disorders.

The result raises an age-old question: do we think through words, or do words primarily serve to communicate thoughts constructed elsewhere? The similarity between syntax and logic has long fueled the idea of ​​a single mechanism: both break down complex structures into simpler elements and recombine them according to hierarchical rules. The data from the group coordinated by Evelina Fedorenko, associate professor of brain and cognitive sciences at MIT, instead indicate a division of cognitive laborThe areas that process words, syntax, and meaning are not necessary to infer a rule or verify a conclusion.

The scope is practical. It touches on the evaluation of theaphasia, the judgment on those who communicate with difficulty and the comparison between brain and large language models. For the Research, the methodological point is clear: apparently intertwined capabilities may depend on different neural infrastructures.

Two experimental paths to isolate reasoning

The first part of the study observed neurotypical adults during various tasks inside a scanner. In the open version of the manuscript, the sample includes: 29 participants; subsets of at least 17 people performed each of the three logic tests. Seventeen tackled the induction task, 23 the verbal syllogisms, and 17 the geometric matrices. The researchers also located, individual by individual, both the language network both the so-called multiple demand network, a frontoparietal system associated with complex problem solving and executive control.

In the inductive test, volunteers were shown lists of input and output digits. The task was to discover the hidden transformation: reverse the order, eliminate values ​​above a threshold, or apply another operation. Once the hypothesis was formulated, the rule had to be applied to new examples. The other tests asked them to evaluate syllogisms constructed with "if-then" propositions or to identify relationships between figures. This variety is crucial because it reduces the risk of confusing logic with a single format, whether verbal, numerical, or visual.

Functional magnetic resonance imaging measures changes in the BOLD signal, which is related to blood oxygenation, and does not directly read thoughts. The team therefore used independent localizers: sentences compared with sequences of nonwords to identify linguistic regions, and spatial memory tests of varying difficulty to identify the multiple-demand network. The design tests whether previously defined circuits systematically respond when the inference demand increases.

Induction and deduction activate different brain systems

The central data is negative, but informative: the language network is not recruited neither in induction nor in deduction. In the open manuscript, deductive contrasts do not produce a significant response in language areas; in induction, a small effect appears, while the contrast used to localize linguistic processing is over four times stronger. Therefore, simply presenting a problem with words is not enough for the inferential computation to occur in the same circuits that comprise the sentence.

The separation becomes more complex when observing the multiple demand networkThis network responds when participants must derive a rule from examples, but it does not show the same involvement in deduction. Based on the reported results, deductive reasoning appears to rely on regions distinct from both the linguistic and generalist executive systems. Therefore, a single center of logic does not emerge: induction and deduction They could distribute the work across partially different architectures. This is an important indication, but it still doesn't amount to a complete map of the mental representations used for reasoning.

The second line of evidence involves two men with profound aphasia, aged 78 and 50, both with extensive lesions in the inferior frontal and temporal areas of the left hemisphere. Their results were compared with those of 40 neurotypical participants of comparable age. Despite showing severe impairments in grammatical comprehension and production, the two patients solved induction problems and logical matrices at a level compatible with the control group. In some cases, they expressed the discovered rule through gestures or sketches.

“This result overturns the idea that symbolic rules require language,”

observes Hope Kean, first author of the study and postdoctoral researcher.

The clinical sample is necessarily very small and does not allow for generalizations about all forms of aphasia. However, the lesion method offers information that neuroimaging alone cannot provide: if a system is severely damaged and a function remains available, the hypothesis that that system is indispensable becomes less plausible. The convergence between absence of activation in healthy participants and performance preserved in patients, it therefore strengthens the interpretation, while leaving open questions about individual variability and recovery after brain damage.

Aphasia, cognitive abilities and the risk of misjudgments

The most immediate implications concern clinical practice and social perception. A person with aphasia may struggle to find words, understand a sentence, or produce grammatical structure, yet continue to play chess, solve sudoku, manage money, and make complex decisions. Confusing a communication difficulties with a reduction in intelligence it can alter functional diagnoses, rehabilitation paths, family relationships and possibilities for autonomy.

“Language difficulties do not indicate how intelligent or capable a person is,”

says Evelina Fedorenko.

The principle extends, with due differences, also to those who stutter, those with developmental linguistic conditions, or those who use a non-native language. For institutions, businesses, and educational systems, research suggests separating them more conceptual competence, decision-making ability e expressive fluidityInterviews, exams, and service interfaces that are highly dependent on verbal speed can underestimate people who reason correctly but communicate in less conventional ways. This is an interpretative consequence, not an automatic clinical prescription.

Rehabilitation can also benefit from more granular assessmentsMaintaining reasoning does not make the linguistic damage mild, but it allows therapists and caregivers to use diagrams, gestures, numbers and visual representations without considering them cognitively inferior substitutes. The contribution of the University College London, through Rosemary Varley's group specializing in acquired language disorders, shows the value of combining basic neuroscience and clinical observation.

From language models to more modular AI systems

The relationship with theartificial intelligence It's inevitable, but it requires caution. Systems like ChatGPT and Claude are trained primarily on large text collections and produce text, yet they can simulate different forms of reasoning. The human brain, according to this work, organizes language and logic into distinguishable networks. The difference doesn't prove that language models don't reason, nor does it prescribe an alternative architecture; rather, it offers an experimental criterion for comparing. fluent verbal production e abstract inference.

For those who develop systems machine learning algorithm , the most concrete consequence concerns benchmark and evaluationA model can formulate a persuasive explanation without consistently applying a rule, or find the correct solution without being able to verbalize it well. Benchmarks and audit procedures should therefore distinguish the accuracy of the conclusion, the stability of the procedure, the ability to generalize a rule, and the quality of the explanation. Neuroscientific research does not directly provide these tests, but it questions the use of eloquence as a sufficient indicator of competence.

Looking ahead, architectures that combine linguistic components with planning, memory, computation, or symbolic manipulation modules could also be studied in light of this biological separation. However, it would be inappropriate to turn the analogy into a recipe: artificial neural networks and the brain do not necessarily share the same constraints, and the MIT work describes human organization, not the industrial efficiency of a computational solution.

The research published by PNAS, accessible through the page of the MIT News and scientific paper, however, narrows the scope of hypotheses: speaking and reasoning can cooperate, but they do not coincide. For neuroscience, this means mapping the geography of inferences; for medicine and society, it means preventing the way a person expresses a thought from being mistaken for the measure of that thought itself; for AI, it means rigorously separating verbal performance from logical ability.

Here are three insights that might interest you:

Mapping the Brain: Adrian Wanner's Swiss Revolution
There is more parallel "traffic" in the human brain than in the animal brain
Mankind one step away from… telepathy thanks to Alter Ego