Research
I study how online harms are measured, detected, and understood in real systems. My work focuses on scams, fraud, unsafe interactions, coordinated abuse, information integrity, and human trust in AI-mediated safety decisions. Across these projects, I build human-centered and computational methods for identifying risky behavior, evaluating safety interventions, and understanding when automated systems succeed or fail in messy platform settings.
Detecting harm in real-world conditions
Online safety problems rarely appear in clean, isolated forms. They unfold through conversations, images, videos, platform affordances, social context, and adversarial behavior. My work studies these risks using ecologically valid data and methods designed around how people actually communicate online. With youth as collaborators, I built datasets and models for detecting unsafe interactions in Instagram direct messages, grounded in young people’s own experiences and threat models (CHI 2022, Honorable Mention; CSCW 2023, Impact Recognition Award). This research showed that safety signals are often multimodal: images, video, interaction patterns, and context matter, not just text. That matters as platforms adopt end-to-end encryption, harmful behavior becomes more coded and evasive, and text-only moderation becomes less reliable.
Abuse, fraud, and information integrity at scale
Online harm often moves across accounts, communities, and platforms. My research studies how abuse, fraud, and manipulation appear in platform data, and how to measure these harms reliably. I have examined how banned users migrate after deplatforming (WebSci 2021), how state-sponsored trolls operate across Reddit and Twitter (IEEE S&P 2022; RAID 2024), and how survey fraud can distort research findings, platform measurement, and downstream AI evaluation (SOUPS 2026). This line of work connects online safety with measurement integrity: detecting harmful behavior is not enough if the underlying data, labels, metrics, or evaluation pipelines are unreliable. I develop methods for measuring abuse, identifying adversarial behavior, and evaluating whether safety interventions reduce harm in practice.
Trustworthy AI as a human-centered problem
As AI systems play a larger role in safety, fraud, and integrity decisions, the central question is not only whether they are accurate, but whether people can understand, contest, and appropriately rely on them. My current work at TRAILS examines how people interpret AI-generated content, authenticity and provenance signals, and visible AI use in shared and workplace settings. I study how people separate related but distinct judgments, such as misinformation, offensiveness, fairness, politeness, trust, and personal risk. Better safety systems should preserve these distinctions instead of collapsing them into a single score, especially when those scores shape platform enforcement, user trust, research validity, or other high-stakes decisions.