Artificial intelligence (AI) systems are showing unexpected strength in language analysis, performing tasks once thought to be the sole domain of human experts. In recent testing, a large language model developed by OpenAI demonstrated reasoning abilities comparable to those of trained linguists—suggesting that AI can now handle aspects of language with precision similar to expert human understanding.
The findings come from a study led by Gašper Beguš of the University of California, Berkeley, along with linguist Maksymilian Dąbkowski and Ryan Rhodes of Rutgers University.
The team examined whether large language models (LLMs) could go beyond simple text generation to actually analyze language structure—a task requiring skills historically limited to humans.
AI model matches linguists in complex syntax tasks
The researchers tested several models, focusing on syntax, ambiguity, and phonology. One model, referred to as o1, stood out. It parsed sentence structures, identified multiple interpretations of ambiguous phrases, and even handled complex concepts like recursion—where phrases are nested within other phrases.
In one test, o1 broke down a sentence into its layered structure and then added a deeper level of recursion, mirroring the analytical methods used by graduate linguistics students.
Beguš said the model’s performance challenged existing beliefs about what artificial intelligence can achieve in linguistic analysis.
David Mortensen, a computational linguist at Carnegie Mellon University not involved in the study, noted that the results suggest language models are moving closer to actual language understanding, not just surface-level prediction.
Ambiguity and phonology reveal depth of AI’s language ability
In another task, the model addressed ambiguity in a sentence like “Rowan fed his pet chicken,” producing two different syntactic structures—one interpreting the chicken as a pet, the other as food. Tom McCoy, a linguist at Yale University, said this kind of reasoning is typically difficult for machines because it requires commonsense knowledge.
To ensure the models weren’t relying on memorized data, the team created a set of new languages containing 40 invented words each. In analyzing these unfamiliar phonological patterns, o1 correctly identified a rule involving breathy vowels following specific consonants. Mortensen, reviewing the outcome, said the model’s accurate analysis was far more advanced than expected.
The results raise broader questions about whether AI’s performance will continue to scale with more data and computing power—or whether some aspects of human language are deeply tied to biology.

