In a groundbreaking development, large language models (LLMs) have demonstrated the ability to analyze language with a level of sophistication previously thought to be uniquely human. A recent study led by linguists Gašper Beguš of the University of California, Berkeley, Maksymilian Dąbkowski, and Ryan Rhodes of Rutgers University shows that one model, OpenAI’s o1, can perform complex linguistic analysis on par with a trained human expert. The research challenges long-held views, including those of Noam Chomsky, who argued that language understanding requires deep, rule-based reasoning that cannot be achieved through data alone. The team designed a rigorous test to assess whether LLMs could go beyond pattern recognition and actually reason about language structure. To prevent the models from simply recalling information from their training data, the researchers created entirely new linguistic tasks. These included parsing sentences with complex recursive structures—where phrases are embedded within other phrases—especially the most difficult form, center embedding. For example, the sentence “The astronomy the ancients we revere studied was not separate from astrology” requires careful syntactic dissection. The o1 model not only correctly analyzed the structure but also extended it with additional layers of recursion, showing a deep grasp of hierarchical sentence organization. The model also handled ambiguity effectively. In the sentence “Rowan fed his pet chicken,” humans use context and common sense to distinguish between a live animal and a meal. The o1 model generated two separate syntactic trees, each corresponding to a different interpretation—something that has historically been difficult for AI systems. In another test, the researchers created 30 fictional mini-languages with invented phonological rules. The model was asked to infer how sounds were produced and combined. For one language, o1 correctly deduced that “a vowel becomes breathy when preceded by a voiced obstruent consonant”—a rule that was never explicitly taught and not present in any training data. Experts outside the study, such as Yale’s Tom McCoy and Carnegie Mellon’s David Mortensen, called the findings significant. They suggest that LLMs may possess a form of metalinguistic ability—the capacity to think about language structure itself—rather than merely mimicking it. Still, researchers caution that while these models can analyze language with human-like precision, they do not yet generate original insights or demonstrate creativity. Their performance stems from vast training data and computational power, not true understanding. As Mortensen noted, current models are optimized for predicting the next word, not for generalizing from limited examples. Nonetheless, the results mark a turning point. They show that what was once considered a uniquely human trait—deep linguistic reasoning—can be replicated, at least in part, by AI. As Beguš put it, the findings represent a steady erosion of the idea that human language is beyond the reach of artificial systems. The line between human and machine language ability is blurring, raising profound questions about the future of intelligence and cognition.