AI Language Understanding: A Neuroscientist's View

Can AI Truly Understand Language?

Veena D. Dwivedi, Director – Centre for Neuroscience; Professor - Psychology | Neuroscience, Brock University, receives funding from the Canada Foundation for Innovation, the Social Sciences and Humanities Research Council of Canada, and Brock University. Brock University provides funding as a member of The Conversation CA-FR. Brock University provides funding as a member of The Conversation CA.

We use language to understand our experiences. The emergence of AI like ChatGPT raises questions about the definition of “meaning.” Some say AI tools “understand” what they do. Geoffrey Hinton stated neural networks understand natural language well and process language better than the Chomskyan school of linguistics.

Chomskyan linguistics involves Noam Chomsky’s theories about language and its development and proposes a universal grammar innate in humans that allows language acquisition from birth.

Having researched language understanding since the 1990s, including over 20 years of neuroscience studies, I disagree with the idea that AI can “understand.”

Written Text vs. Natural Language

People often confuse text with natural language. Written text relates to language but isn’t the same. The same language can have different visual symbols, like Hindi and Urdu, or Serbian and Croatian. Written text isn’t “language.”

Machine Learning and Context

Linguistic communication happens face-to-face, with environmental context, tone, eye contact, and emotional expressions. Understanding involves more than comprehending words. Even babies understand context cues.

For example, the sentence “I’m pregnant” has different meanings depending on who says it and in what context. The recipient ascribes different meanings to the same sentence.

Research shows an individual’s emotional state alters brainwave patterns when processing sentence meaning. Our brains always have emotional context, so computer code responding to text doesn’t capture human understanding.

AI neural networks are computer algorithms, not biological brain networks. Confusing “flight” (birds) with “flight” (airlines) can cause misunderstandings.

Chomskyan Linguistics and AI

Neural networks don’t process language better than Chomskyan linguistics theories. This field assumes languages can be understood via grammatical systems and a universal grammar.

Chomsky researched syntactic theory but didn’t conduct experiments on the psychological or neural bases of language comprehension. His linguistics ideas are silent on sentence processing and understanding mechanisms.

The Chomskyan school of linguistics asks how infants learn language easily, barring deficits. With at least 7,000 languages, the human brain must be ready to comprehend and learn language at birth. Chomsky posited an innate module for language learning.

The brain must be ready to understand language from birth. There are examples of language specialization in infants, but the neural mechanisms are unknown.

Misusing scientific terms and conflating AI with human understanding can lead to dangerous consequences.