I’m Dr. Kevin McAlister — associate teaching professor in the Department of Data and Decision Sciences at Emory University and senior faculty fellow with Emory’s Center for AI Learning. I teach data science principles, probabilistic machine learning, and the mathematical foundations of modern generative AI at every level: to students with nothing but algebra, through the lens of critical thinking and systems understanding, and to advanced students with the mathematics, probability theory, and computer science to understand modern generative AI from first principles. I direct a continuum of applied AI and data science programs — Emory’s largest experiential data science programs — where students of all skill levels learn to ship real systems for real organizations. Beyond teaching, I build AI tools for people who aren’t AI people, and I develop educational policy and content that puts AI at the forefront of learning and understanding.
AI was built to make lives better, and I believe it can — if we build it with care. But a tool for the people, by the people, has to be understood by the people, or there is no holding its builders accountable when what they make drifts from what communities value. My work is about making that understanding possible, for people from all walks of life.
Teaching — one thesis, two altitudes.
The same probabilistic foundations at two altitudes — graduate generative ML with interactive walkthroughs and its full lecture set, and a first-year intro taught with algebra and a coin-flip game.
In Production — systems in daily use.
Census pipelines for grant writers, curatorial judgment encoded at a museum, and courses designed around AI — each with a deliberate answer to what the model is trusted with, and what it never is.
AI Education Programs — a continuum.
Three stages from data science elementary school to high school: 700+ students and 150+ real projects for civic and corporate partners.
Working On — the lab notebook.
Sports-analytics latent-variable models, exact single-shot generation, and the influence horizon of language models — research in flight, each stated at its true stage.
About — the short version.
Probabilistic methods, an Amazon experimentation patent, and the interdisciplinary toolkit behind the teaching and the systems.
My Philosophy
AI is the most human exercise ever undertaken. Underneath the engineering, building these systems has forced us to take questions that have eluded philosophers and factory workers alike — what does it mean to think? to speak? to be creative? — and make them concrete enough for a pile of electrified rocks to work with. The images we see are local arrangements of color and line. The language we produce comes from years of hearing how words go together. Music is a time series of frequencies that people, for some deeply human reason, decided sounded good. Generative AI works because humanity’s collective experience turned out to be compressible — and making the result actually feel human is a problem the technologists cannot solve alone. That’s not a diminishment of the field. It’s the most interesting thing about it.
Taking that seriously means being honest about what these systems are: probabilistic machine learning with good marketing. Nearly everything in applied statistics is fancy 1-nearest-neighbor — find the thing in the training data closest to what you’re trying to do, and bet that the best guess is what we saw before. GPT is that idea applied to enough text to learn very good conditional probabilities. Vision models are that idea applied to enough images to learn how small patches of the world fit together. And every piece of the machinery encodes a human choice: architectures are priors, losses are priors, training data is a prior, benchmarks are priors. The question is never whether the priors are there — they always are — but whose knowledge got encoded, and with what care. That framing explains the failures as readily as the successes. We live in a biased world, so models trained on it will be biased without deliberate intervention. These systems learn averages, which is why generated faces drift toward samey. A style that millions of humans recognize can’t simply be deleted from a model, because the model learned it the same way we did. None of this requires a PhD to see. It requires a handful of data principles and the habit of asking: where did this come from, and what did we lose when we compressed it?
I care about who gets to ask those questions because of where I come from. I’m a first-generation college student from a working-class family, and when my parents ask me about AI, their questions aren’t abstract: whether nursing work gets automated out from under people in the last third of their careers, whether the factory lines do, whether data centers swallow the lakes, beaches, and mountains of North Carolina — from the center of the state, each is about a two-hour drive — that they’re counting on enjoying in retirement. I think we’re near a fork in the road — AI built by and for people, in service of making lives genuinely easier, or AI built by the few, serving the few, leaving everyone else to absorb the costs. Which path we take depends less on the models than on how many people understand them well enough to shape what they’re for. My job is to grow that number.
And that job is changing underneath everyone who teaches. What is a teacher for, when every learner has a tireless tutor that can explain anything, at any level, on demand? My honest answer: explanation — the thing we’ve treated as the core of teaching for centuries — is about to become cheap, and that’s mostly good news, because explanation was never the hard part. The hard part is deciding what’s worth explaining, in what order, and how you’d know if it worked. A tutor that can explain anything will cheerfully explain things in the wrong sequence, build intuitions that collapse under the next concept, and tell a learner they understand when they’ve only learned to nod along. What doesn’t get cheap is curriculum: the load-bearing structure underneath the explanations. When I built my graduate generative models course, the design decision that mattered wasn’t any individual lecture — it was the sequencing that makes diffusion models feel inevitable if you understood week three, and the two-notebook structure that forces every concept to survive contact with running code. So my position on AI-era teaching is a design principle: let the model adapt the path; the human architects the terrain — deciding what’s true, what’s foundational, what a learner needs to struggle with rather than be handed, and whether anyone actually learned rather than merely finished. That’s what I build. The pages here are the evidence.