Evidence from Formal Logical Reasoning Reveals that the Language of Thought is not Natural Language

By combining fMRI studies in healthy adults with behavioral assessments of individuals with severe aphasia, this research demonstrates that logical reasoning operates independently of natural language, engaging distinct neural networks rather than linguistic representations.

The Gist

Imagine your brain is a bustling city with two very different neighborhoods: The Language District and The Logic District.

For a long time, many scientists and philosophers believed these two neighborhoods were actually the same place. They thought that to do complex math, solve a puzzle, or figure out a logical argument, your brain had to use the same "machinery" it uses to speak and understand sentences. In other words, they thought we think in words.

But a new study by a team of researchers from MIT, UCL, and Berkeley says: Not so fast.

Using high-tech brain scanners and studying patients with severe language damage, they found that these two neighborhoods are actually separate. You can lose your ability to speak and understand language, but your ability to do logic remains perfectly intact.

Here is the breakdown of their discovery, using some simple analogies:

1. The Two Neighborhoods

Think of your brain like a city with specialized zones:

• The Language District: This is where you process grammar, vocabulary, and the meaning of words. It's like the city's library and post office combined.
• The Logic District: This is where you solve puzzles, spot patterns, and figure out "If A, then B." It's like the city's engineering and planning department.

The old theory said: "To build a house (logic), you must use the library's blueprints (language)."
The new study says: "No, the engineers have their own blueprints. They don't need the library to do their job."

2. The Experiment: Scanning the City

The researchers did two main things to prove this:

A. The "Healthy Brain" Scan (fMRI)
They asked healthy people to do two types of tasks while inside an MRI machine (which takes pictures of brain activity):

• Task 1 (Language): Read sentences.
• Task 2 (Logic): Solve pattern puzzles.
  • Inductive Reasoning: "Here is a list of numbers: 2, 4, 6. What comes next? Figure out the rule."
  • Deductive Reasoning: "If it rains, the grass is wet. It is not wet. Therefore, it did not rain."

The Result: When the healthy people did the logic tasks, their Language District stayed quiet. It was like a library that is closed for the day. Instead, a different part of the brain (the "Multiple Demand Network," or the city's general-purpose problem-solving team) lit up like a Christmas tree.

B. The "Damaged City" Test (Aphasia Patients)
This is the most dramatic part. They studied two men who had suffered massive strokes that destroyed their Language District.

• These men had severe aphasia. They could barely speak, couldn't understand complex sentences, and struggled with grammar. If you asked them, "Is the cat on the mat?" they might not understand the question.
• The Test: The researchers gave them logic puzzles.
  • The Puzzle: "Here are some shapes. One doesn't fit. Which one is it?" or "Here are numbers. What's the pattern?"
• The Result: Despite being unable to speak or understand language, these men solved the logic puzzles perfectly. Their "Logic District" was working just fine, even though their "Library" was in ruins.

3. The "Jabberwocky" Twist

In one part of the study, they used sentences that looked like English but were nonsense (like "If the tep is ag, then the wug is zib").

• If you needed real language to do logic, your brain should struggle with these nonsense words.
• But the brain didn't care. It treated the nonsense words just like variables in a math equation ($x$ and $y$). It ignored the "meaning" and just looked at the structure. This proves the brain is using a structural, logical system, not a language system.

4. Why Does This Matter?

This changes how we understand the human mind.

• We aren't just "talking machines": Our ability to reason isn't just a side effect of learning to speak. We have a built-in "Logic Engine" that exists independently of our vocabulary.
• The "Language of Thought" is different: Philosophers used to say we have a "Language of Thought." This study suggests that this internal language isn't English, Chinese, or Spanish. It's a different kind of code—more like computer code or a mathematical formula—that runs on a different part of the brain.
• Resilience: It shows that if you lose your ability to communicate, you don't lose your ability to think. Your mind is still sharp, even if your voice is gone.

The Bottom Line

Think of your brain as a smartphone.

• Language is like the Keyboard App. It's great for typing messages and reading emails.
• Logic is like the Calculator App.

For years, people thought you couldn't use the Calculator without the Keyboard. But this study shows that even if you smash the Keyboard (or if the Keyboard is broken), the Calculator still works perfectly. You can still do the math; you just can't type it out.

Conclusion: We don't think in words. We think in a special, logical code that happens to be separate from the language we use to talk to each other.

Technical Summary

1. Problem Statement

The paper addresses a longstanding debate in cognitive science and philosophy regarding the relationship between natural language and logical reasoning.

• The Hypothesis: Many scholars (e.g., Fodor, Chomsky, Carruthers) have argued for the "Language of Thought" (LOT) hypothesis, positing that complex thought, including formal logical reasoning (inductive and deductive), relies on linguistic representations. This view suggests that the brain uses the same syntactic and hierarchical mechanisms for thinking as it does for processing natural language.
• The Counter-Hypothesis: Alternative views suggest that logical reasoning is a distinct cognitive system, potentially relying on non-linguistic representations (e.g., mental models, abstract symbols, or probabilistic programs).
• The Gap: While previous studies have shown separability between language and other domains (like math or music), the specific relationship between abstract formal logical reasoning and the language network remains unclear. Prior fMRI studies have yielded mixed results, often confounding logical difficulty with general cognitive effort or failing to include appropriate language control tasks. Furthermore, behavioral evidence from aphasia patients has been limited by the severity of impairment required to test the hypothesis.

2. Methodology

The authors employed a two-pronged approach combining functional magnetic resonance imaging (fMRI) in healthy adults and behavioral testing in individuals with severe aphasia.

Study 1: fMRI in Healthy Adults

• Participants: 29 native English speakers with canonical left-lateralized language networks.
• Tasks:
  1. Inductive Reasoning: Participants inferred rules governing transformations between input and output number lists (e.g., [5, 7] → [7, 5, 7]). Rules involved mathematical, list, and structural operations. The task contrasted the Induction phase (hypothesis generation) against an Application phase (rule execution).
  2. Deductive Reasoning (Syllogistic): Participants judged the validity of conclusions based on two premises. The task varied by difficulty (Modus Tollens vs. Modus Ponens) and stimulus type (Real words vs. Non-words).
  3. Deductive Reasoning (Matrix): A non-verbal task where participants identified an outlier in a matrix of geometric shapes (fMRI version) or completed a missing piece (behavioral version).
• Localizers:
  • Language Network: Defined using a standard localizer (Sentences > Non-word lists) to identify functionally specific regions of interest (fROIs) in the left inferior frontal and temporal cortices.
  • Multiple Demand (MD) Network: Defined using a spatial working memory task to identify domain-general executive regions.
• Analysis: Linear mixed-effects models were used to compare BOLD signal responses in the language network and MD network across logic contrasts.

Study 2: Behavioral Testing in Aphasia

• Participants: Two individuals (S.A. and G.S.) with profound, chronic aphasia resulting from extensive lesions to the left peri-Sylvian cortex (affecting Broca's and Wernicke's areas). Both exhibited severe grammatical impairments (agrammatism) and near-chance performance on syntactic comprehension tasks.
• Tasks:
  • Inductive Reasoning: Same rule-learning task as Study 1 (paper-based).
  • Deductive Reasoning: Matrix Reasoning subtest from the WASI-II (Wechsler Abbreviated Scale of Intelligence).
• Comparison: Performance was compared against age-matched neurotypical controls using Crawford-Howell single-case tests.

3. Key Contributions

• Neural Dissociation: The study provides robust evidence that the language network (specifically sensitive to syntactic structure) is not engaged during formal logical reasoning tasks, neither for induction nor deduction.
• Double Dissociation of Reasoning Types: The authors demonstrate that different forms of logic recruit distinct neural systems:
  • Inductive Reasoning recruits the Multiple Demand (MD) network (domain-general executive control).
  • Deductive Reasoning (specifically complex syllogisms) recruits a distinct network separate from both the language network and the MD network.
• Aphasia Evidence: The study establishes that individuals with severe, permanent damage to the language system retain intact logical reasoning abilities, challenging the necessity of linguistic representations for thought.
• Methodological Rigor: The study addresses limitations of prior work by using precise fMRI localizers, non-verbal logic tasks (matrix reasoning), and patients with profound (rather than mild) language deficits.

4. Results

fMRI Findings (Study 1)

• Language Network Response: The language network showed a strong response to the language localizer (Sentences > Non-words, $p < 0.001$). However, it showed no significant response to logical reasoning contrasts:
  • Inductive reasoning: Minimal response ($p < 0.05$ at network level, but effect size was negligible compared to language; induction response was at or below the level of reading non-words).
  • Deductive reasoning (Syllogistic): No significant response for Modus Tollens > Modus Ponens ($p > 0.1$).
  • Deductive reasoning (Matrix): No significant response ($p > 0.1$).
  • Note: The lack of response held true even when using real words vs. non-words in syllogisms, suggesting the brain treats these as variables rather than linguistic stimuli during logic tasks.
• MD Network Response:
  • Inductive Reasoning: Robustly engaged the MD network (Induction > Application, $p < 0.001$).
  • Deductive Reasoning: Did not engage the MD network ($p > 0.48$).
• Whole-Brain Search: Deductive reasoning (Modus Tollens > Modus Ponens) activated a distinct set of frontal and parietal areas not overlapping with the language or MD networks.

Behavioral Findings (Study 2)

• Language Impairment: Both patients performed at or near chance on syntactic assessments (e.g., assigning thematic roles in reversible sentences).
• Logical Reasoning Preservation:
  • Inductive Reasoning: S.A. solved 19/25 rules; G.S. solved 39/40 rules. Both performed within the normal range of controls (no significant difference).
  • Matrix Reasoning: S.A. scored 25/30; G.S. scored 26/30. Both scored significantly above the age-matched normative mean (T-scores of 73 and 68, respectively).
• Conclusion: Severe linguistic impairment does not preclude high-level logical reasoning.

5. Significance

• Refutation of the "Language of Thought" Hypothesis: The findings strongly suggest that natural language is neither utilized nor required for abstract logical reasoning in the adult human brain. The brain does not rely on linguistic syntactic structures to perform formal logic.
• Modularity of Cognition: The study reinforces the view that the human brain is composed of specialized, domain-specific systems. The language network is dedicated to linguistic computation, while logical reasoning relies on distinct representational formats (potentially abstract symbols, mental models, or probabilistic programs).
• Implications for AI and Development:
  • AI: The results challenge the assumption that large language models (LLMs) achieve reasoning capabilities simply by mastering natural language. True logical reasoning may require architectures distinct from those used for language generation.
  • Development: While language may facilitate the acquisition of logical concepts in childhood, the underlying mechanism of reasoning appears to be independent of language.
• Clinical Relevance: The preservation of logic in aphasia suggests that rehabilitation strategies for cognitive deficits should not assume that restoring language is a prerequisite for restoring logical reasoning.

In summary, the paper provides converging neuroimaging and neuropsychological evidence that logical reasoning is a distinct cognitive faculty supported by neural systems separate from the language network, fundamentally decoupling the "Language of Thought" from natural language.