Current Projects

Improving Human Detection of Deepfakes: I am investigating how people detect AI-generated images and what factors make them miss the subtle artifacts these images contain, under supervision of Matt Groh and Jessica Hullman at Northwestern. My research aims to develop tools and techniques to enhance human ability to distinguish between real and synthetic images. In a world where AI-generated images are increasingly realistic, understanding how people perceive them is crucial for combating misinformation and maintaining trust in visual media.

Think you can distinguish real photos from AI-generated images? Try our DetectFakes website!

Mechanistic Interpretability of AI-Image Detection: I am studying what open-weight vision-language models know about AI-generated images and whether they use it.

LLM Coach for Cash-Transfer Recipients: I led Northwestern’s evaluation of a WhatsApp LLM coach that GiveDirectly deployed to cash-transfer recipients in Rwanda. I helped design a randomized comparison of two coaching prompts, built an arm-blind LLM-as-judge rubric for topics, coaching style, and plan formation, and analyzed more than 1,500 real voice and text conversation turns.

Decision-Calibrated Explanations: I contributed the human judgment data behind HD-RL, a reinforcement learning method that fine-tunes an open-weight LLM to write explanations that improve human decisions, using a reward model trained on real human choices.

Past Projects

Conformal Prediction for AI-Advised Decisions: Advised by Jessica Hullman, I researched enhancing AI-advised decision-making by accurately quantifying prediction uncertainty of deep neural networks. I explored the effectiveness of conformal prediction sets as an alternative to traditional uncertainty measures, measuring how they impact human decisions in real-world image labeling scenarios. This work received a Best Paper Honorable Mention at CHI 2024.

Prenatal Stress Reduction: In collaboration with the Center for Advancing Safety of Machine Intelligence (CASMI), advised by Maia Jacobs, I worked on co-designing patient-facing machine learning for prenatal stress reduction.