What if the most vaunted AI breakthroughs of the decade are chasing a target that’s fundamentally wrong? For all the breathless predictions by tech leaders of imminent superintelligence, neuroscience and cognitive science are converging on a stark conclusion: language isn’t intelligence and scaling LLMs won’t bridge that gap.
The core architecture behind systems like ChatGPT, Claude, and Gemini is optimized for one thing: predicting the next token in a sequence of text. This statistical machinery excels at producing coherent language, but it’s fundamentally a model of linguistic form, not of thought. As Evelina Fedorenko, Steven Piantadosi, and Edward Gibson put it in Nature, “Language is primarily a tool for communication rather than thought.” fMRI studies reveal that the neural networks for reasoning, abstraction, and causal inference are distinct from those for language processing. Patients with severe language impairments can still solve math problems, infer others’ intentions, and reason logically-evidence that cognition can operate independently of linguistic ability.
Developmental neuroscience supports this dissociation. Infants, long before they utter their first words, do sophisticated hypothesis testing about the physical and social world. Work using awake infant fMRI shows that early cognition is grounded in perception, action, and multimodal sensory integration, not in text-like symbolic streams. This embodied, non-linguistic learning scaffolds later language acquisition, but it is not reducible to it. By contrast, LLMs are trained almost exclusively on text, deprived of the sensory and motor grounding that gives shape to human intelligence.
This text-only paradigm increasingly reveals its limits when tested adversarially. In the medical domain, the benchmark mARC-QA is designed to disrupt familiar statistical cues deliberately in order to provoke the Einstellung effect, which is rigid reliance on learned patterns of responses. Even state-of-the-art models such as GPT‑4o, Gemini 1.5‑Pro, and DeepSeek‑V3 achieved less than 53% accuracy, which is far behind the 66% average attained by human physicians, while often making high-confidence errors. Such shortcomings represent more than just data gaps but rather architectural brittleness in reasoning and generalization when low-probability or out-of-distribution events occur.
Yann LeCun has been blunt: current autoregressive LLMs “lack essential capabilities like understanding the physical world, persistent memory, reasoning, and planning.” The alternative he has presented envisions Joint Embedding Predictive Architectures that learn from sensory data with a focus on abstract and causal representations to enable models to simulate and interact with the world. That seems closely related to what Fei‑Fei Li had in mind with her call for “world models” that maintain spatial consistency, adhere to physical laws, and integrate multimodal input in support of grounded reasoning. Such systems would be closer to how people build internal models of reality-a prerequisite for robust planning and flexible problem-solving.
The gap, from the perspective of cognitive science, is not just about sensory grounding but about the architecture of intelligence itself. Human cognition is a composite of specialized subsystems-perceptual processing, motor control, memory consolidation, social reasoning-coordinated to act in dynamic environments. The recent attempt by Yoshua Bengio, Eric Schmidt, and Gary Marcus at defining AGI corresponds to the “cognitive versatility and proficiency of a well-educated adult” and implicitly acknowledges that no single scaling law could unify these abilities inside a monolithic language model.
But even if engineers were able to integrate a battery of domain competencies, there’s another hurdle: dissatisfaction-driven creativity. Philosophers like Richard Rorty, following Thomas Kuhn’s theory of paradigm shift, argue that the most interesting scientific insights follow from rejecting the dominant conceptual vocabulary in any field. Einstein developed his theory of relativity well before experimental verification, with motivation nothing more urgent than a feeling that Newtonian mechanics was incomplete.
LLMs have no native incentive, however, to question their training data; they are, fundamentally, statistics about it. Deprived of means for both generating and testing new frames beyond what’s captured by learned distributions, they threaten to become “dead-metaphor machines” repositories of recombined common sense, incapable of the conceptual leaps that define knowledge.
The technical way forward likely lies with hybrid architectures that will integrate grounded world models, adaptive planning, and metacognitive calibration with language capabilities. Such work would need to move beyond token prediction toward systems that can store persistent state, simulate counterfactuals, and learn from embodied interaction. Until then, the conflation of linguistic fluency with general intelligence will remain a category error, which neuroscience, developmental studies, and adversarial AI testing are continually dismantling.
