Artificial intelligence (AI) has transformed dramatically over the past few decades, but for Professor Susan McRoy, its defining feature has always been change. A longtime faculty member and chair of the Department of Computer Science, McRoy has seen large language models (LLMs) evolve into today’s massive, data-driven models that can quickly judge what is meant in context.
LLMs are a specific kind of AI that is language- and text-focused, compared to other kinds of AI used for image classification, fraud detection, predictive analytics, and decision making. Both LLM and other kinds of AI encompass machine learning, where computers learn from data and identify patterns without being explicitly programmed.
McRoy has been interested in communication between people and computers on matters of health, where nuances in the language can pose a particular obstacle to understanding. She is also interested in developing explanations for machine learning models which help people trust them and verify that they are working properly.
In this Q&A, McRoy reflects on the field’s rapid transformation, how students are encountering AI today, and why language remains one of the most intriguing challenges in computing. (Interested in related courses to get started? See the listing at the end.)
AI seems to have spread everywhere in what seems like a short time. What happened?
The recent rise of AI came from a convergence of rapid computing abilities because of graphical processing unit (GPU) advances and the explosion of internet data – conditions that allowed researchers to revisit old problems with new architectures.
AI is very good at memorizing and reproducing common patterns. If a task repeats often, AI can capture it – similar to how programmers save useful pieces of code so they can reuse them instead of starting from scratch. But its strength is not universal; it depends heavily on training data.
How do you explain the difference between computer science and AI?
Computer science is the broader field. AI sits inside it and has always been the more experimental path. Some applications require strong guarantees that systems work correctly; AI hasn’t always offered that provability. As it becomes more capable, it still inherits that experimental nature.
How did you first get interested in computer science?
What drew me in was the idea of getting computers to do things people do – an inherently evolving challenge. When I was starting college, my father found an article predicting that computer science would be the future. It was a field that felt completely new at the time—there were no high-school classes then, and computers were very primitive.
Why did you choose natural language processing (NLP) as your research area?
Language provides the most direct window into how people think. Early AI researchers used logic and text to model reasoning. My work focused on interactive communication – how people misunderstand each other, how systems can detect inconsistencies, and how to repair misunderstandings in real time. These challenges are still with us today.
How do today’s large language models relate to that early work?
Early systems combined grammar, meaning, and context using hand-written rules and tiny training sets. Modern models do something similar but automatically and at massive scale. They’re trained on everything online – both good and bad – which shapes their output.
How should students use AI?
They can use it to support—not replace – critical thinking. Ask AI to help you evaluate ideas or identify related work so you don’t duplicate existing research. Let it broaden your understanding, fuel new questions – not do the thinking for you.
How has the field changed during your career?
The problems haven’t changed much, but the methods have changed completely. Change is the defining feature of AI, so faculty who’ve been here a long time have had to continually adapt.
Three introductory AI or ML courses for CEAS students
- COMPSCI 290, “Introductory Topics in Computer Science: Trending and Trustworthy Artificial Intelligence.” No prerequisites, but students must be in the college.
- COMPSCI 411, “Machine Learning and Applications.” Freshman must have at least one programming class.
- COMPSCI 422, Introduction to Artificial Intelligence.” Prerequisites COMPSCI 317 and 351.