Effects of Large Model Biases in IR

This lecture analyzes biases, focusing on large language and multimedia models. First, it discusses different types of biases that affect the model training pipeline, including human, data-related, and alignment-related biases. Second, it exemplifies their effects in information-seeking tasks now powered by large models, notably in high-impact domains such as politics, health, and finance. Finally, it introduces a few recent works that aim to reduce bias and thus improve fairness.