Seminar Details

In this talk, I demonstrate how the interplay between EconCS theory and AI can mitigate misalignment across three stages of the data-to-decision pipeline: data collection, data evaluation, and decision-making. I present two recent projects that design manipulation-robust evaluators to (i) achieve incentive alignment by eliciting genuine, effortful human feedback under general reward schemes, and (ii) ensure quality alignment by assessing the desired information value in datasets that are contaminated by AI-generated content, without relying on ground truth. The talk concludes with a high-level discussion of policy alignment as a broader research direction that my recent work speaks to, outlining how robust human and AI decision policies can be designed under strategically generated or noisy AI-synthetic data.