Scientists are investigating what makes humans unique, and language has long been considered a key element. Aristotle once said that man is “the animal that has language.” Although large language models such as ChatGPT can imitate ordinary speech, researchers have wondered whether there are aspects of human language that have no parallels in animals or artificial systems. Recently, they have focused on the models’ ability to reason about language itself. Some linguists, including Noam Chomsky, argued in a 2023 article in The New York Times that AI models cannot perform sophisticated analysis because correct explanations of language are complex and cannot be learned from large amounts of data alone.
This view was challenged by a study by Gašper Beguš, a linguist at the University of California, Berkeley, Maksymilian Dąbkowski, who recently earned his PhD in linguistics at Berkeley, and Ryan Rhodes of Rutgers University. The scientists subjected several large language models (LLMs) to a series of linguistic tests, including generalizing the rules of an invented language. Most models failed to parse linguistic rules the way humans do, but one model demonstrated impressive abilities that exceeded expectations. It was able to analyze language like a college linguistics student—drawing sentence diagrams, resolving multiple meanings, and working with complex features such as recursion.
Infinite Complexity in Tests
One challenge in testing language models is ensuring that they are not merely repeating learned answers. The models are trained on vast amounts of text, including internet content in dozens of languages and linguistics textbooks. To avoid this, Beguš and his colleagues created a four-part linguistic test. Three parts involved analyzing specially constructed sentences using tree diagrams, first introduced by Chomsky in his 1957 book Syntactic Structures. These diagrams break sentences down into noun and verb phrases, and then further into nouns, verbs, adjectives, adverbs, prepositions, and conjunctions.
One part of the test focused on recursion, which is the ability to embed phrases within phrases. For example, the simple sentence “The sky is blue” can be expanded to “Ana said that the sky is blue.” This process can continue indefinitely, for example, “Maria wondered whether Sam knows that Omar heard Jane say that the sky is blue.” Chomsky and others describe recursion as a defining feature of human language, making it possible to generate an infinite number of sentences from a finite vocabulary and set of rules. So far, there is no convincing evidence that animals use recursion in a sophisticated way.
Recursion can occur at the beginning or end of a sentence, but the most difficult form is center embedding, as in the transition from “the cat died” to “the cat the dog bit died.” The test included 30 original sentences with difficult examples of recursion, such as “Astronomy, which ancient people studied, was not separate from astrology.” Using a syntactic tree, the OpenAI o1 model was able to identify the structure as: Astronomy [which the ancients [whom we worship] studied] was not separate from astrology. It then added another layer of recursion: Astronomy [which the ancients [whom we worship [who lived in countries that we respect]] studied] was not separate from astrology.
What Do You Mean? Distinguishing Meanings
Tom McCoy, a computational linguist at Yale University who was not involved in the research, was surprised by the performance of the o1 model, especially its ability to recognize ambiguity, which has traditionally been difficult for computational models. People use common sense to rule out unlikely meanings, but computers lack this kind of knowledge. For example, the sentence “Rowan fed his pet chicken” can mean either that Rowan fed his pet hen or that he fed his pet chicken meat. The o1 model correctly created two different syntactic trees for the two meanings.
The research also included experiments with phonology, the study of sound patterns and the organization of phonemes, the smallest units of sound. People learn phonological rules through practice, without explicit instruction. In English, for example, adding “s” to words ending in “g” produces a “z” sound, as in “dogs,” while after “t” it has the standard “s” sound, as in “cats.”
The team created 30 new mini-languages, each containing 40 invented words, such as θalp, ʃebre, ði̤zṳ, ga̤rbo̤nda̤, ʒi̤zṳðe̤jo. The models were tasked with analyzing phonological processes. For one language, the o1 model correctly determined that a vowel becomes breathy when immediately preceded by a consonant that is voiced and obstruent—a sound produced by restricting airflow, such as “t” in “top.” These languages were new, so the model could not have encountered them during training.
Uniquely Human or Not?
David Mortensen, a computational linguist at Carnegie Mellon University, described the results as surprising because models traditionally only predict the next word, which differs from deep human understanding. According to him, this work challenges the claim that LLMs do not use real language. The question is how far the models will go—will they improve simply by increasing computing power and data, or are some properties of language evolutionarily limited to humans?
The results show that models can perform sophisticated analysis, but they have not yet invented anything original or uncovered new insights about language. If improvement is only a matter of scale, Beguš believes that models will surpass us in language skills. Mortensen notes that current models are limited by being trained to predict tokens, but he expects that over time they will generalize better from smaller amounts of data and do so more creatively.
Research is gradually stripping away traits that were once considered exclusively human. According to Beguš, this suggests that we are less unique than we thought. McCoy emphasizes the importance of understanding where models succeed and where they fail, because society increasingly depends on them. Linguistic analysis is an ideal test for assessing the extent to which they reason like humans.
Sources: quantamagazine.org and ieeexplore.ieee.org
