The Neuroscience Case Against LLM Intelligence: Why Language ≠ Thought

November 29, 2025 • 8 min read

The Core Argument: Language Is Not Thinking

Benjamin Riley, founder of Cognitive Resonance, recently published a compelling essay arguing that the AI industry's bet on large language models is fundamentally misguided. The core insight comes from neuroscience: human thinking is largely independent of human language.

This isn't philosophical speculation—it's backed by decades of brain imaging research. Evelina Fedorenko, Associate Professor at MIT's Department of Brain and Cognitive Sciences, has published extensive work demonstrating that language and thought are distinct in the human brain.

Perhaps more striking: studies of people who lost their language abilities (through stroke or other brain damage) showed that their ability to think remained largely unimpaired. They could still solve math problems, follow nonverbal instructions, and understand others' emotions.

The Mathematical Ceiling on AI Creativity

A new study by David H. Cropley, Professor of Engineering Innovation at the University of South Australia, adds mathematical weight to the argument. Published in the Journal of Creative Behavior, the research calculates a fundamental upper limit on LLM creativity.

The mechanism is straightforward: LLMs predict the next most probable token based on training data. High-probability selections yield effective but predictable output. Low-probability selections might be novel but often become nonsensical. You can't maximize both simultaneously.

Yann LeCun's Rebellion: From Meta to "World Models"

The skepticism isn't limited to academics. Yann LeCun—Turing Award winner, deep learning pioneer, and until recently Meta's Chief AI Scientist—has been one of the most vocal critics of the LLM paradigm. He's now leaving Meta to start a company focused on "world models."

LeCun's core argument: LLMs are trained on text, but text represents only a tiny fraction of the information that even a young child processes about the world.

LeCun offers a simple thought experiment that illustrates the limitation:

Any human can do this effortlessly. An LLM can write about rotating cubes, but it has no internal representation of what that actually means in physical space. It lacks what LeCun calls a "world model"—an internal representation of physical structure, dynamics, and cause-and-effect relationships.

What Are "World Models" and Why Do They Matter?

World models represent a fundamentally different approach to AI. Instead of predicting the next word in a sequence, they aim to build internal representations of how the physical world works—like humans (and even cats) do naturally.

1. Sensory Data Ingestion: The system takes in streams of perceptual data (visual, tactile, etc.)
2. Latent State Learning: It builds compressed internal variables that capture what's actually happening in the world
3. Predictive Dynamics: It learns how that internal state will evolve when the agent or environment acts
4. Planning Module: A separate component uses this machinery to plan and choose actions

This is more like how biological intelligence works. A house cat has never read a physics textbook, but it can plan highly complex actions based on an intuitive understanding of how objects, gravity, and space work. LeCun argues we need AI systems that learn from sensory experience, not just language.

The Counter-Arguments: Why LLMs Might Still Matter

Not everyone agrees with the "LLMs are a dead end" thesis. Some counter-arguments worth considering:

Emergent Capabilities: LLMs have repeatedly demonstrated capabilities that weren't explicitly trained—suggesting that scaling might unlock unexpected abilities.

Multimodal Extensions: Models like GPT-4V, Gemini, and Claude already process images and other modalities, not just text. The line between "language models" and "world models" may blur.

Practical Utility: Even if LLMs aren't "intelligent" in some philosophical sense, they're enormously useful for coding, writing, analysis, and many other tasks. The question of "true intelligence" may be academic.

Hybrid Approaches: Even LeCun acknowledges that LLMs will likely persist as modules for turning thoughts into words. The future might be LLMs + world models working together.

The Bottom Line

The neuroscience is clear: language and thought are separate processes in the human brain. The math is clear: LLMs face a fundamental creativity ceiling due to their probabilistic architecture. The industry dynamics are clear: one of AI's founding figures is betting his career on an alternative approach.

This doesn't mean LLMs are useless—they're extraordinarily capable tools. But the industry's claims about achieving AGI through scaling language models may be, as LeCun puts it, a "dead end."

For those of us building and deploying systems, the practical takeaway is to use LLMs for what they're good at (language tasks, pattern matching, coding assistance) while remaining skeptical of claims that they'll soon be able to reason about the physical world, plan complex actions, or achieve genuine understanding.