CS 378: Introduction to AI Safety

Spring 2027 · Mondays, 3:00–6:00 PM

Instructor: Dori Litvak · dorlit at utexas dot edu

Office hours: Tuesdays, 3:00–4:00 PM

Overview

As AI systems become extremely capable and are deployed in high-stakes settings, what can go wrong? Can we anticipate it, and what technical challenges do we face?

Introduction to AI Safety introduces the technical, ethical, and governance challenges of building and evaluating safe AI systems.

We'll start with the foundations: how large language models are trained, how they game their objectives, absorb bias from data, and fail in the real world. From there, we'll cover the main tools the field uses: red-teaming and jailbreaks, interpretability, formal verification, scalable oversight, and evaluations at scale. We finish with the bigger picture, including governance and policy, economic incentives, frontier risks, and open questions about model psychology, constitution, persona, and welfare.

We will learn by reading research papers, examining case studies, gaining hands-on technical experience, and practicing academic writing. The final project includes original writing, participating in a peer-review process, and presenting at the UT AI Safety Symposium.