Interpretative Interfaces: Designing for AI-Mediated Reading Practices and the Knowledge Commons

Explanation isn't enough for understanding: Simply telling users how AI systems work doesn't enable them to actually engage with or question those systems. Scientists using AI for literature reviews particularly need hands-on ways to probe how models process information, not just passive explanations of their outputs.

Users should be able to directly manipulate AI's "thinking": Rather than treating AI as a black box, the research proposes letting users interact with the model's intermediate representations—essentially following how individual words (tokens) change meaning as the AI processes context, similar to how readers have historically annotated texts.

Historical reading practices offer design lessons: Techniques like marginalia, glosses, and annotations have helped readers engage critically with texts for centuries. These same principles can be adapted to help people interact meaningfully with AI systems rather than passively accepting their outputs.

Current AI interfaces fail knowledge workers: Scientists and other professionals who rely on AI for research tasks lack tools to verify how the model arrived at its conclusions or to intervene when something seems wrong—a significant gap for fields where accuracy and traceability matter.

Interactive engagement protects the knowledge commons: By enabling users to understand and annotate how AI transforms information, interpretative interfaces could help preserve critical thinking and prevent unchecked AI-generated content from degrading shared knowledge resources.