ChatGPT can write essays, explain calculus, and generate computer code in a fraction of the time it takes a human. What it can't replicate is the kind of casual conversation shared between friends on a front porch. 

For Ryan Ka Yau Lai, assistant professor in the University of Georgia's Franklin College of Arts and Sciences Department of Linguistics, the reason is straightforward. 

"At their core, AI models are trained to produce text, not to converse," Lai said. 

That insight is shaping both Lai's research and the classroom. With seed grant funding from Franklin's I-Cubed program, Lai is helping launch “LING 2150--Demystifying Conversational AI: Language, Mind, and Society,”a new course that pairs linguistics with computer science. Students will learn how people communicate, how large language models work, and why the two are not the same. A companion computer science course will explore how LLMs are built and trained. 

Lai's latest research shows that users move fluidly between treating AI as a conversational partner and as a tool. Those shifting expectations reveal both the strengths and the limits of today's conversational AI. 

One conversation in the study begins with a user greeting ChatGPT with a simple "Hey" before making a request. After ChatGPT responds, the user repeats almost the entire prompt, changing only a single word to generate a different result. When the response requires correction, the user drops the greeting altogether and simply tells ChatGPT to revise its previous answer. The exchange illustrates how users shift between conversational communication and button-pushing.  

"What we're seeing is the same question that has shaped AI since its beginnings,” said Lai. “Do we engage a computer as an 'it'—something to control and use—or as a 'you'—someone we can address and converse with? Today's language models are more fluent than ever, but that question remains unanswered." 

The same pattern appeared throughout the conversations Lai analyzed. At times, users treated ChatGPT like a conversational partner, referring to earlier exchanges and asking follow-up questions. At other times, they treated it like a tool, issuing new commands with no connection to the previous response. Those interactions suggest people approach AI with expectations drawn from both conversation and computer use. 

Those expectations collide with the way large language models are built. In conversation, people interrupt, clarify, negotiate meaning, and reshape what they say as they receive feedback. AI systems generate complete responses before the interaction continues. 

For Lai, the research demonstrates why linguistics belongs in conversations about artificial intelligence. 

"Conversation is not about the mere transmission of context-free information," he said. "Language use is about intersubjective alignment between subjects, where actions, knowledge, and perspectives are shared, compared, and contested." 

As conversational AI becomes part of education, health care, and everyday life, Lai believes future AI professionals will need more than engineering expertise. They will need to understand how people communicate. 

“If AI is to truly serve people, humanistic and social-scientific perspectives cannot be  afterthoughts,” said Lai. “They must be woven into every stage of AI development.”