Why Can’t AI Truly Understand Human Language? Chomsky and Moro Explain the Biological Roots of Grammar in The Secrets of Words

Is Human Language Hardwired in the Brain? What Chomsky and Moro’s The Secrets of Words Teaches Us About How Every Child Learns Grammar

Noam Chomsky and Andrea Moro make a compelling case that language is biologically encoded in every human brain — and no AI model or big data system can replicate it. Discover what experiments with “impossible languages,” universal grammar, and neuroscience reveal about what truly separates human speech from machine imitation.

Keep reading to find out why two of the world’s leading linguists believe the key to understanding human language lies not in datasets or algorithms — but deep within the biological architecture of every human brain.

Recommendation

Noam Chomsky has famously argued that human language is akin to a snowflake: Just as water molecules acting in accordance with the rules of temperature can result in a host of complex patterns and shapes, language develops according to inborn rules, yet can unfold in unknowably complex ways. In this short but dense text, Chomsky — “the father of modern linguistics” — and linguist and neuroscientist Andrea Moro converse about the past, present, and future of linguistics and the study of the human brain. Though the ideas are intriguing, readers unfamiliar with the field may find parts of the text hard to parse. It also (unsurprisingly) ignores the controversial status of some of Dr. Chomsky’s ideas.

Take-Aways

• Early linguistics theorists assumed languages were based on infinitely variable, arbitrary sets of rules, which infants learned by listening and practicing until they became habits.
• Chomsky’s new paradigm, “generative grammar,” proposed that all languages depend on a universal set of rules.
• All human languages display an essential unity in that, regardless of the seeming variations in their complexity, children learn them all around the same age.
• Through experiments with “impossible languages,” Andrea Moro and others have found evidence to support Chomsky’s theory.
• Linguistic structures are biological in nature — machines can imitate syntactic constructions, but they cannot comprehend the actual mechanisms underlying language.
• The modern enthusiasm for big data — the belief that every mystery can be resolved by feeding sufficient data into a sufficiently powerful computer — is misplaced; true advances take the form of theories that explain things.
• The present and future challenge in linguistics is to investigate the relationship between the formal structures of syntax and the neurobiological structures of the brain.

Summary

Early linguistics theorists assumed languages were based on infinitely variable, arbitrary sets of rules, which infants learned by listening and practicing until they became habits.

In the pre-Chomsky world, when linguistics was attempting to establish itself as a new area of scientific inquiry, most people took it for granted that languages were essentially arbitrary assemblages of oral and/or written symbols and that the rules governing them could vary without limits. Linguists concerned themselves with mapping out the physical structure of languages — the sounds and characters — and trusted that a general theory of language would emerge from such analysis. That did not prove to be the case.

During this time, radical behaviorism dominated the field of psychology, and language was seen as simply one more complex behavior that would eventually be resolved into B.F. Skinnerian stimulus-response (S-R) connections. Language was assumed to be, in the words of Leonard Bloomfield, “a matter of training and habit.”

“Others believed the same thing: Children are trained with thousands of examples, millions of examples, and somehow the habit is formed and they know what to say next. If they produce or understand something new, it’s by ‘analogy.”

In the late 1940s, Karl Lashley, one of the pioneers of neuroscience, demonstrated that S-R analysis could not possibly account for simple behaviors — like how a horse gallops — much less for more complex ones — like how a person plays music. But it would be years before psychology began to pay attention to these insights.

Such impasses have occurred at numerous inflection points in the history of science. The neo-scholastic world of Newton and Galileo was stymied by the phenomenon of gravity since — so the thinking went — objects could not act on one another over a distance. The physics of 100 years ago (pre-quantum) could not account for observations in the field of chemistry. In both cases, progress in human understanding would not depend on more data but on a new theoretical model that would explain what people were seeing.

Chomsky’s new paradigm, “generative grammar,” proposed that all languages depend on a universal set of rules.

Chomsky’s 1957 book Syntactic Structures provided that new theoretical model by proposing that language as a whole is a process rooted in the biology of the brain. In his formulation, the key to understanding language as a phenomenon was syntax, the hierarchical rules governing how words can be arranged in sentences.

Nearly every human infant learns to speak the language of its culture and to speak it with consistent grammar. But to suggest that infants master the rules of a language simply by listening to the examples they hear is simply not reasonable. The data is too sparse and inconsistent, and the edifice to be constructed is too complex, to be explained as simply imitation with no biological contribution.

“Children learn grammar the way they learn to walk or digest (and no child learns how to digest by means of imitation).”

The imitation model offers a parsimonious explanation for the miraculous accomplishment that is human language. Human brains are biologically prepared to build hierarchical structures of symbols, recombining those symbolic structures as needed to generate infinite numbers of statements from a finite set of symbols.

All human languages display an essential unity in that, regardless of the seeming variations in their complexity, children learn them all around the same age.

The old understanding of languages was taxonomic and morphological: It grouped them into families of derivation and cataloged the differences in how people formed words and put them in order to form sentences. Studying the seemingly vast differences between languages that grew from different roots, researchers failed to recognize their underlying sameness.

“Despite all superficial and apparent differences, every human being in essence speaks the same language, the same way that every human being in essence has the same face.”

This does not mean the differences between the many languages are insignificant or unimportant. The sameness is in the underlying biological template, which limits the ways they can vary. The differences are in which parts of that template are selected for use in any given language and which parts become forgotten, a process that appears to involve some “pruning” of synapses in the vast neural network of the brain.

What researchers have learned is that children are all born with the same innate set of capacities for language learning — capacities that are equally applicable to any human language. Their ability to learn the language they grew up hearing does not depend on what language their biological parents spoke. Also, the age at which they master the grammar of their language does not depend on that language.

Through experiments with “impossible languages,” Andrea Moro and others have found evidence to support Chomsky’s theory.

Andrea Moro and others measured brain activity in research subjects as they learned one of two types of simplified kinds of language: “impossible” and “possible.” They based the rules for impossible language strictly on word order — for example, a rule that says to turn a statement into a question the speaker must reverse word order, so “America is beautiful” becomes “Beautiful is America” and is understood as an interrogative. Possible language followed the normal hierarchical rules that human children use when learning language: “America is beautiful” becomes “Is America beautiful?” in the interrogative.

The researchers conducted two experiments. In the first, subjects learned “micro-versions” of a real language, like Italian, that contained both impossible and possible rules; in the second, they asked subjects to identify impossible and possible rules in a pseudo-language composed of non-real primary words and real linking words (like “the”). Moro and his team found that the subjects in the experiments could distinguish grammatical from ungrammatical sentences in both instances — that is, they learned the possible and impossible rules, even in the pseudo-languages where sentences were rife with made-up words, like, “The gulks janidged the brals.”

However, the two groups showed different patterns of brain activity when identifying possible and impossible rules. When subjects were learning possible rules, they showed increased activity in the areas of their brains related to language learning (part of Broca’s area, in the frontal lobe). When they were learning impossible rules, they showed decreased activity in the language areas and increased activity in the areas associated with solving puzzles.

“What [young children] ignore is the linear order of words…What they pay attention to is the structures they are creating in their minds.”

These experiments suggest that word order is far less important in the communication process than was once believed. This finding is consistent with the results of other research on language acquisition in early childhood. What infants learn is not word order but grammatical structures (“generative grammar”) that can be strung together, nested, and otherwise recombined recursively without limits.

Linguistic structures are biological in nature — machines can imitate syntactic constructions, but they cannot comprehend the actual mechanisms underlying language.

The earlier view of the diversity of human languages led to the assumption that linguistic rules must be as arbitrary as game rules and that, by setting down those rules with sufficient precision, it might someday be possible to build computers that think and talk in human language.

What is meant by language, however, is not a human invention that people knowingly construct but a biological and evolutionary one. No other species, even among our closest evolutionary cousins, has shown a corresponding capacity, and it is difficult to imagine its acquisition by a mechanical device.

“Many experiments have confirmed a hypothesis that Descartes had already asserted at an intuitive level, namely, that syntax constitutes the fundamental distinction between the language of humans and that of all other living beings: We are the only ones capable of recombining a bounded set of discrete elements (broadly, words) to generate a potentially infinite set of expressions (broadly, sentences).”

There is now a tremendous enthusiasm for artificial intelligence, neural nets, and so forth. But simulating human language use through a computer program will never be the same as what a human infant does when it activates innate neural capabilities to learn to talk.

The modern enthusiasm for big data — the belief that every mystery can be resolved by feeding sufficient data into a sufficiently powerful computer — is misplaced; true advances take the form of theories that explain things.

Linguists before Chomsky, like many researchers today, believed that the answers to the problems they studied were discoverable by the magic of statistical analysis. Not only did they attempt to deduce the regularities of structure in languages in that way, but they also argued that young children figured out the rules of their own languages simply by listening to other native speakers.

The problem with their assumptions is that the rules are too complex, and the data — both for infant language learners and for adult researchers — is too sparse.

“Capturing the syntax of human languages in full by analyzing an immense number of sentences would be like capturing the fact that the sun is fixed and we rotate around it by taking thousands of trillions of pictures of the sun through the window. Scientific research simply doesn’t develop like that.”

Children learn to understand and speak their native languages spontaneously — and to follow rules of grammar that they may never learn to articulate — because their brains are genetically programmed for that kind of learning. Researchers in linguistic science cannot take advantage of any such pre-programming, and their analyses were leading nowhere before the publication of Syntactic Structures.

Like other scientific revolutions in the past — such as universal gravitation and quantum physics — Chomsky’s recognition of the biological contribution to language learning advanced the science of linguistics by providing a new paradigm that would explain many puzzling phenomena, bring together previously separated fields of inquiry, and change the kinds of questions scientists were asking.

The present and future challenge in linguistics is to investigate the relationship between the formal structures of syntax and the neurobiological structures of the brain.

Chomsky’s work on syntactic structures transformed the understanding of human language and provided the conceptual framework necessary to study its underlying structure.

“The discovery of this amazing link between language structure and the brain is so revolutionary that it can be expressed by reversing the 2,000-year-old traditional perspective and arriving at the surprising conclusion that it’s flesh that became logos, not vice versa.”

Thanks to modern neuroimaging techniques, scientists now know where language happens in the brain. What remains to be explored, probably awaiting the next new paradigm, is the connection between syntactic structures — the arrangements of words and phrases within a sentence — and neuronal structures — how neurons receive, process, and transmit information.

There are other challenges, of course. Researchers still do not understand how languages change and although some of these mysteries will probably be solved by future scientists, others may remain forever just beyond humanity’s grasp.

About the Authors

Theoretical linguist Noam Chomsky’s 1957 book Syntactic Structures provided the paradigm for the modern science of linguistics and remains near the top of every list of the most influential nonfiction works ever published. Andrea Moro is a professor of general linguistics at the Institute for Advanced Study IUSS Pavia, Italy, where he researches and writes about syntax and neurolinguistics.