Mitigating Harm in Online Information Systems

Abstract

Online information systems increasingly mediate what individuals read, encounter, believe, and discuss. Platforms such as X, formerly Twitter, shape access to news and public debate through sophisticated recommendation algorithms. More recently, large language models (LLMs) have become a foundation for systems that help users access and synthesize knowledge. However, these systems can produce significant harms. Their algorithmic design may contribute to homogeneous exposure, political polarization, one-sided news coverage, and unreliable generated information. As a result, those systems can restrict individuals’ exposure to diverse content and perspectives while reinforcing existing biases. In this thesis, we aim to deepen our understanding of these challenges by examining the design choices through which information systems become narrow, polarizing, incomplete, or unreliable. We formalize the underlying computational problems from complex real-world challenges, analyze the design choices that produce harmful outcomes, and develop efficient algorithms to measure, mitigate, or address them.

The thesis examines how harms can be represented explicitly in the objectives of the systems that produce them. Our contributions span three connected settings: recommendation, online media, and language-model-based information access. For recommendation, we develop a differentiable, model-agnostic objective for jointly optimizing relevance and set-level diversity during training. For online media, we introduce an opinion-dynamics model that links aggregate timeline composition to network-level polarization, together with a scalable convex optimization method for reducing the resulting disagreement--polarization index. We further show that source diversity is an insufficient proxy for viewpoint diversity in news aggregation and formulate calibrated viewpoint coverage as a constrained selection problem. For language models, we adopt context-controlled evaluations to show that misleading context can induce confident errors that token-level uncertainty does not reliably detect. Hidden-state probes provide a stronger signal for identifying these unreliable responses. Across these settings, the thesis argues that mitigating informational harm requires objectives and evaluations that directly represent the relevant societal property, rather than relying solely on proxies such as predicted relevance. 

All experiments in this thesis are conducted on real-world datasets under rigorous evaluation settings, and the proposed methods are compared against multiple baselines.Implementations of the proposed methods, together with the datasets released as part of this work, are publicly available.

Link to DiVA