DOF: New AI tool finds blind spots in document collections

A new research paper from Youdi Li, accepted at the CHI 2026 Workshop on Tools for Thought, challenges how AI helps us explore large text collections. While current tools like summarization and topic modeling optimize for coverage—highlighting main themes—they inevitably push minority viewpoints, edge cases, and unexpected findings out of view. This blind spot hinders discovery, which often depends on noticing what doesn't fit. Li proposes an alternative objective: blind-spot discovery, where the goal is to surface the very content that coverage methods suppress, letting users judge its significance.

The proposed system, DOF (Discovery-Oriented Faceting), achieves this through three design goals: organizing documents into categories with explicit boundaries, ranking categories by distinctiveness rather than size, and supporting iterative refinement. In a comparative evaluation against coverage-based ranking across four domains, DOF surfaced fundamentally different content—promoting specialized categories that traditional methods buried. This approach offers a complementary mode of support for researchers and analysts seeking to uncover gaps, minority positions, or unexpected patterns in large document collections.