No—not doomed, but language-only learning may encounter a ceiling.
The MIT study does not show that intelligence cannot be learned from language. It shows that, in humans, the neural machinery that performs formal reasoning is separable from the machinery specialized for understanding and producing natural language. People with severe damage to the language network could still solve certain nonverbal logical problems, while healthy participants’ language regions did not respond systematically to increasing reasoning difficulty.
That finding has three different implications for AI.
1. It challenges the strongest LLM hypothesis
The strongest hypothesis would be:
Scale language prediction far enough, and the whole of intelligence must emerge from it.
The MIT finding makes that hypothesis less biologically plausible. The human brain apparently does not obtain all reasoning by running an unusually powerful internal language processor. It has distinguishable systems for language and at least some forms of logical cognition.
Therefore, an AI trained only to predict linguistic tokens may be learning intelligence through a narrower aperture than humans do.
Language contains enormous compressed traces of reality—science, mathematics, history, social experience and human explanations. A model can infer much about the world from those traces. But linguistic data are still reports about experience, not experience itself.
A text-only model has never directly:
- watched an object continue to exist after disappearing behind another;
- moved through three-dimensional space;
- discovered by action that a cup falls when released;
- distinguished something dangerous from harmless through bodily consequences;
- negotiated an unpredictable physical environment.
It learns these things from what other beings have said about them.
In that sense, text-only learning resembles someone who has read the entire medical library but has never encountered a patient.
2. But language training can produce nonlinguistic internal computation
The contrary inference would also be too strong:
Because human reasoning is not performed by the language network, an AI trained on language cannot learn reasoning.
That does not follow.
The training material and the internal mechanism are not the same thing. A model trained through language may develop latent representations of quantity, space, causation, programs and abstract relations that are not themselves ordinary sentences.
For example, learning millions of linguistic descriptions of chess could allow a system to construct an internal model of the chessboard. The input remains language, but the learned representation need not remain linguistic.
Thus, the critical question is not:
Does an AI receive words?
It is:
What internal structures does it build from those words?
A language model might use natural language as the entrance and eventually construct something resembling a nonlinguistic language of thought. The MIT paper does not rule this out. Indeed, its central point—that structured thought need not be natural language—could equally apply to the hidden representations inside an artificial model.
3. The likely ceiling concerns grounding, not formal logic alone
Curiously, formal logic may not be the greatest weakness of language-trained AI. Logic, mathematics and code are highly represented in symbolic data. AI systems can receive explicit correctness signals and generate many synthetic exercises.
The harder limitation is likely to concern grounding:
- Which features of reality matter?
- What persists when appearances change?
- What action is possible here?
- What does an unexpected event mean?
- When is a verbal description misleading?
- How does an agent revise its model after reality resists it?
Language can describe resistance, but embodied interaction supplies actual resistance.
This is where the distinction between a world described and a world encountered becomes decisive.
A purely linguistic system may possess an immense map of human descriptions without possessing the same kind of causal contact with the territory. That does not make it unintelligent, but its intelligence may remain characteristically different: encyclopedic, combinatorial and representational, yet comparatively fragile when the environment departs from its textual precedents.
Language may be a bootstrap, not a prison
We should distinguish four possibilities:
|
AI architecture |
Likely capacity |
|
Text-only prediction |
Rich linguistic and conceptual intelligence, but weakly grounded |
|
Language plus images/audio/video |
Better perceptual world modelling |
|
Multimodal AI plus tools and persistent memory |
Can test propositions and accumulate consequences |
|
Embodied, agentic AI acting in an environment |
Can learn through intervention, error and resistance |
The historical importance of LLMs may therefore be that language serves as an extraordinary bootstrap for intelligence, not that language alone constitutes intelligence.
Once linguistic learning has produced sufficiently powerful conceptual machinery, AI can be connected to perception, simulation, tools, robotics, experiments and long-term interaction. Its later intelligence need not remain confined to the medium through which it first learned.
Humans also begin under constraints. We do not directly perceive economic systems, galaxies or quantum fields; we extend cognition through language, diagrams, instruments and institutions. AI may travel in the reverse direction:
Humans move from embodied life into language and abstraction.
AI begins in language and may have to move outward into a world.
Where I think the real limitation lies
The MIT paper does not imply that AI intelligence is doomed to be limited. All intelligences are limited by architecture, sensors, learning histories and environments—including ours.
It suggests something more precise:
Natural language is neither identical with intelligence nor guaranteed to be a sufficient training environment for every form of intelligence.
The future of AI will probably require differentiation analogous—not identical—to the human brain:
- systems for language;
- systems for spatial and causal modelling;
- systems for planning;
- memory extending across time;
- perception and action loops;
- metacognitive systems that detect uncertainty and error.
In other words, intelligence may require not one indefinitely enlarged language faculty, but an ecology of specialized capacities capable of temporarily assembling around a problem.
And here the finding touches our distinction between the algorithmic and analogical self. An LLM can become enormously proficient at operating over symbolically preserved regularities. But analogical intelligence also arises from bodily situation, perceptual wholes, accidental encounters, historical residue and the resistance of things that refuse to fit the available sentence.
So I would put it this way:
Language may carry AI remarkably far, perhaps farther than anyone expected. But language alone cannot guarantee a world.
The ceiling is not necessarily imposed by words. It is imposed when words never have to answer to anything outside words.
