New research shows AI can analyze language with skills rivaling human experts. Credits: Jernej Furman from Slovenia, CC BY 2.0, via Wikimedia Commons
Artificial intelligence (AI) is now showing remarkable skill in language analysis — an area once believed to be the exclusive territory of trained linguists. A recent study reveals that an advanced AI system can handle complex language tasks with an accuracy comparable to human experts.
The research, led by Gašper Beguš of the University of California, Berkeley, along with Maksymilian Dąbkowski and Ryan Rhodes from Rutgers University, tested whether large language models could go beyond generating sentences to actually understanding how language works. The findings show that AI can now perform in-depth linguistic analysis, challenging earlier assumptions about the limitations of machines in understanding human language.
AI matches human reasoning in sentence analysis
The study focused on how well AI could process parts of language, such as sentence structure, word ambiguity, and sound patterns. One model, named o1 (pronounced “oh-one”), delivered standout results. It broke down sentences, identified multiple possible meanings, and even managed tasks that involved layers of meaning — something human language experts typically learn through years of study.
In one key example, o1 took a sentence and restructured it by adding another layer of meaning within the original sentence. This task, known as recursion, is commonly taught at the graduate level in linguistics. The AI’s ability to handle it closely mirrored how a trained linguist would approach the challenge.
These findings surprised many in the research community. David Mortensen, a computational linguist from Carnegie Mellon University who was not part of the study, noted that the results show AI is starting to understand how language actually works — not just how to predict the next word in a sentence.
AI understands ambiguity and sound rules in new languages
The model also handled sentences that could be interpreted in more than one way. For example, in the sentence, “Rowan fed his pet chicken,” the AI correctly identified two meanings: One where the chicken is a pet and another where it could be food. This kind of analysis often requires common sense — something machines have traditionally struggled to apply.
To avoid the risk that the AI was simply repeating patterns it had seen during training, the researchers created new, fictional languages. Each one had 40 made-up words. The AI still managed to discover consistent rules within these languages, including one where a certain type of vowel followed specific consonants. According to Mortensen, this kind of reasoning was much more advanced than what was previously expected from AI.
The results raise important questions about how far artificial intelligence can go in mastering human language. As these models continue to improve, researchers are now asking whether future breakthroughs will rely only on more data and faster computers — or if there are limits tied to the human brain that machines may never cross.