My research examines how computational systems can better support human reasoning, decision-making, and social interaction online. I focus particularly on the design and evaluation of personalized AI systems that adapt to individuals’ goals, values, and contexts. Much of this work lies at the intersection of human-computer interaction, social computing, and human-AI interaction.
A central theme in my work is developing computational methods to model human preferences and values, and embed them into societal algorithms, including social media curation and moderation algorithms and generative AI. To evaluate these systems, I conduct user studies and randomized controlled experiments that test how these designs or algorithmic interventions affect user behavior and judgments.
Value Alignment of Social Media Ranking Algorithms
Much of the discussion around “AI alignment” focuses on large language models, but social media ranking algorithms are another powerful class of societal algorithms that shape what people see and engage with every day. Today’s feeds are typically optimized for engagement, which may appear neutral but in practice systematically amplifies certain values over others. My work introduces a framework for value alignment of social media ranking algorithms, showing how feeds can be ranked around human values such as caring, tolerance, or tradition. Because human values are inherently in tension with one another, this method does not impose a fixed set of values. Instead, it allows users to articulate and negotiate the value trade-offs they want reflected in their own feeds.
———–
Personalized Content Transformation for Online Moderation
Most content moderation systems assume that harmful content can be defined the same way for everyone and that moderating it requires hiding or removing it. Our work challenges this assumption by showing that experiences of harm are often deeply personal and contextual, making posts that appear benign to others distressing for some—such as individuals dealing with PTSD or specific phobias.
My work proposes an alternative paradigm for moderation: personalized content transformation. Instead of removing posts entirely, this approach allows users to define what types of content are harmful to them and transforms those elements while preserving the surrounding context. We implemented this idea in a browser extension for Reddit that modifies triggering visual and textual content in real time—for example, by cartoonifying images or reducing visual detail. Our studies show that these transformations help users stay connected to communities and conversations they might otherwise avoid while reducing distress.