Understanding Human Decision Making and Building Aligned AI Systems

About

Understanding how people make decisions is both an important scientific question in its own right and increasingly important in designing and aligning intelligent systems. In this talk, I will present work around these two questions: what can we learn about human decision making and preferences, and how can it help us build AI systems that are better aligned with people?

I will first present preliminary findings on context effects in multiattribute choice (MAC). Rational models of choice assume that preference between two options remains unchanged by the introduction of an alternative. Context effects break these assumptions and show that preferences shift with the added context of a third option. These effects are fragile in the literature, though, frequently disappearing or reversing in experiments. We examine whether the geometry of choice problems explains this variability. In purchasing decisions, things like price tier, relative value advantage of one option over the other, and price difference are geometric parameters that impact outcomes. We capture a large-scale map of multiattribute choice behavior by systematically varying the geometric parameters proposed and explore how they modulate context effects. We further evaluate competing MAC theories and work toward a more unified account of when and why these effects emerge.

I will then present another line of work on using knowledge of human preference to build better AI systems. We find that learning preferences for LLM response styles and routing across model providers can better align system behavior with user preferences and increase utility. And finally, I will show a quick demo of a conversational AI system that I am building to explore how AI can augment human decision making.

Details

Date:

Friday, September 18, 2026

Time:

12:00 PM - 1:00 PM

Location:

CDS 1646

Theme:

AI, Social Science