A growing body of neuroscience research is challenging one of the most fundamental assumptions underlying modern artificial intelligence: that sophisticated reasoning must be tightly coupled to language. One of the most significant contributions to this debate is the paper “Evidence from Formal Logical Reasoning Reveals That the Language of Thought Is Not Natural Language,” by Hope Kean, Alexander Fung, Paris Jaggers, Jason Chen, Joshua S. Rule, Yael Benn, Joshua B. Tenenbaum, Steven T. Piantadosi, Rosemary A. Varley, and Evelina Fedorenko. The paper was published in the Proceedings of the National Academy of Sciences (PNAS) in 2026, one of the world’s leading peer-reviewed scientific journals. The paper is available at: https://www.pnas.org/doi/10.1073/pnas.2520095123.

The research addresses a question that has occupied philosophers, psychologists, neuroscientists, and artificial intelligence researchers for decades: Is natural language the mechanism by which humans perform logical reasoning, or is reasoning performed by an independent cognitive system? The answer has profound implications not only for understanding the human brain but also for designing the next generation of artificial intelligence systems.

Executive Overview

The investigators tested whether the brain’s language network is actually responsible for formal logical reasoning. Their results strongly suggest that it is not.

Using functional MRI studies together with detailed testing of individuals suffering from profound aphasia caused by severe damage to the brain’s language centers, the researchers demonstrated that formal deductive and inductive reasoning remain largely intact even when the language system has been extensively damaged. Rather than depending upon the neural systems responsible for speech and language comprehension, logical reasoning appears to rely upon different neural circuits.

This finding suggests that the human brain separates language from reasoning far more than previously believed.

Research Methodology

The investigation combined two complementary approaches.

The first involved functional magnetic resonance imaging (fMRI) of healthy adult volunteers. Rather than assuming that everyone’s language centers occupy identical locations, the researchers first identified each participant’s individual language network using well-established language localization procedures. They separately identified the brain’s “multiple-demand” network responsible for executive control, attention, and working memory.

Participants then completed several different reasoning tasks while brain activity was measured.

Inductive reasoning tasks required participants to infer hidden rules governing numerical transformations and then apply those rules to new examples.

Deductive reasoning tasks required participants to determine whether conclusions necessarily followed from stated premises, including classical logical structures such as modus ponens and modus tollens.

Additional nonverbal matrix reasoning tasks measured abstract pattern recognition and logical inference independent of ordinary language.

The second component of the study provided particularly powerful evidence.

Two individuals suffering from profound aphasia caused by extensive destruction of the left-hemisphere language system completed largely nonverbal reasoning tests. These patients exhibited severe impairment in speaking, reading, and understanding language, yet retained the ability to solve complex abstract reasoning problems.

Because lesion studies provide causal evidence rather than merely correlational evidence, these patient results significantly strengthened the conclusions drawn from the imaging experiments.

Major Findings

The results were striking.

The brain’s language network responded strongly during ordinary language processing but showed little or no increased activation during either deductive or inductive reasoning.

Instead, inductive reasoning primarily activated executive-control regions associated with the multiple-demand network.

Deductive reasoning produced activity in a partially distinct collection of frontal and parietal regions rather than the classical language areas.

Perhaps even more remarkable were the patient studies.

Despite profound language impairment, both aphasic participants successfully completed demanding induction and matrix reasoning tasks at levels comparable to neurologically healthy control participants.

The evidence indicates that the neural machinery responsible for formal logical reasoning is largely independent of the machinery responsible for natural language.

Understanding Severe Aphasia

One of the most important practical implications concerns patients with severe aphasia.

For many years, profound language impairment has often been incorrectly associated with equally profound intellectual impairment. This research demonstrates that such an assumption can be seriously misleading.

Individuals who have lost much of their ability to speak, read, or understand language may nevertheless retain sophisticated reasoning abilities, problem-solving skills, judgment, pattern recognition, mathematical reasoning, and decision-making capacity.

The inability to communicate effectively should therefore never be interpreted as evidence that higher cognitive functions have been lost.

This conclusion has important implications for clinical assessment, rehabilitation, legal competency evaluations, education, and long-term patient care.

Implications for Artificial Intelligence

Although the research focuses on human neuroscience, its implications for artificial intelligence may ultimately prove even more significant.

Today’s large language models achieve remarkable performance by learning statistical relationships within enormous collections of human language. Their apparent reasoning abilities emerge largely from language training rather than from explicitly designed reasoning systems.

The new neuroscience evidence suggests that human intelligence may operate quite differently.

Rather than using language itself as the reasoning mechanism, the brain appears to use language primarily as a communication interface while abstract reasoning occurs within specialized computational systems.

If this interpretation proves correct, future AI architectures may increasingly separate language understanding from reasoning.

Instead of requiring every reasoning operation to flow through an extremely large language model, future systems may combine several specialized components:

• A relatively compact language model dedicated primarily to communication and understanding human instructions.

• Independent symbolic or mathematical reasoning engines capable of rigorous logical inference.

• Specialized planning systems.

• Knowledge representation systems.

• Memory systems optimized for long-term factual storage.

• Verification modules capable of checking conclusions before presenting them.

Such modular architectures are already beginning to appear in neuro-symbolic AI, retrieval-augmented generation, theorem-proving systems, and multi-agent reasoning frameworks.

The present research provides biological evidence supporting continued exploration of these directions.

Opportunities for Smaller AI Models

One particularly intriguing possibility is the development of substantially smaller models capable of more reliable reasoning.

Current frontier language models require hundreds of billions or even trillions of parameters because language itself carries enormous complexity.

If reasoning can be partially separated from language, future systems might use comparatively compact language models for communication while delegating formal reasoning to highly optimized symbolic engines or specialized neural modules.
 Potential advantages include:

• Lower computational requirements.

• Reduced energy consumption.

• Faster inference.

• Lower deployment costs.

• Improved explainability.

• More reliable logical consistency.

• Easier verification of conclusions.

• Better integration with scientific computation and engineering analysis.

Rather than replacing large language models, these specialized reasoning systems could complement them, allowing each subsystem to perform the task for which it is best suited.

Suggested Future Research

The authors’ work naturally suggests several important directions for future investigation.

Researchers should identify the precise neural circuits responsible for deductive reasoning and determine how they interact with executive-control systems.

Larger studies involving additional aphasia patients with differing lesion patterns would strengthen causal conclusions and identify which forms of reasoning remain preserved across diverse neurological conditions.
 Additional experiments should examine probabilistic reasoning, causal inference, scientific reasoning, mathematical proof, planning, analogical reasoning, and commonsense reasoning to determine whether each depends upon distinct neural systems.

Longitudinal studies of children may clarify whether language is primarily important during cognitive development while mature reasoning later becomes increasingly independent.

From an artificial intelligence perspective, future work should explore biologically inspired architectures that explicitly separate language processing, symbolic reasoning, memory, planning, verification, and decision-making. Such systems may prove more computationally efficient, more interpretable, and more robust than architectures relying exclusively upon increasingly larger language models.

Related Research

Evidence from Formal Logical Reasoning Reveals That the Language of Thought Is Not Natural Language. Hope Kean et al. Published in Proceedings of the National Academy of Sciences (PNAS), 2026. Peer reviewed. This is the principal study reviewed in this article and provides both functional MRI and lesion evidence demonstrating that formal logical reasoning is largely independent of the brain’s language network.

The Boundaries of Language and Thought in Deductive Inference. Martin M. Monti, Lawrence M. Parsons, and Daniel Osherson. Published in Proceedings of the National Academy of Sciences (PNAS), 2009. Peer reviewed. One of the earliest neuroimaging studies demonstrating that logical inference recruits brain regions largely separate from classical language areas.

Neural Dissociation of Algebra and Natural Language. Martin M. Monti and colleagues. Published in PLoS Biology, 2012. Peer reviewed. Demonstrates that algebraic reasoning relies upon neural systems distinct from those supporting ordinary language processing.

Evidence for Cognition Without Grammar from Causal Reasoning and Theory of Mind in an Agrammatic Aphasic Patient. Rosemary Varley and Michael Siegal. Published in Current Biology, 2000. Peer reviewed. An influential lesion study showing that sophisticated reasoning and social cognition can survive severe grammatical impairment.

Dissociating Language and Thought in Human Reasoning. Maria Coetzee and Martin M. Monti. Published in Current Opinion in Behavioral Sciences, 2023. Peer reviewed. A comprehensive review of the evidence supporting the distinction between language processing and formal reasoning in the human brain.

Separating Logic and Language. MIT News, 2026. Not peer reviewed. A high-quality institutional summary explaining the motivation, methodology, and significance of the PNAS research for both neuroscience and artificial intelligence.

Conclusion

For decades, advances in artificial intelligence have often been driven by insights from neuroscience, just as discoveries in AI have provided new ways to think about the human brain. This research continues that tradition by suggesting that one of the brain’s most remarkable capabilities—formal reasoning—may operate largely independently of natural language. If future neuroscience confirms and further refines this model, it could inspire a new generation of AI architectures in which language serves primarily as the interface between humans and machines while specialized reasoning systems perform the underlying logical computation. Such biologically inspired designs could lead to AI systems that are smaller, faster, more energy efficient, more explainable, and more reliable, while advancing our understanding of both natural intelligence and artificial intelligence.