Capability Isn't the Bottleneck Anymore

AI diagnostic tools let novices perform like experts immediately, but reliance on AI for task completion blocks the learning needed to develop actual expertise.

Method: Researchers deployed AI-based cancer diagnostic tools to medical novices in a controlled experiment, measuring both immediate diagnostic performance and skill retention after AI removal. The study isolates whether AI acts purely as a performance crutch or builds transferable diagnostic capability. They tracked diagnostic accuracy with and without AI assistance across multiple sessions, creating a learning curve analysis that separates tool-dependent performance from internalized skill acquisition.

Caveats: Study focuses on cancer diagnosis; skill transfer dynamics may differ in domains with faster feedback loops or lower error costs.

Reflections: What interaction patterns during AI-assisted diagnosis predict better skill retention? · Can interface design choices (e.g., explanation timing, confidence displays) improve learning outcomes without sacrificing immediate performance? · How long does skill retention persist after AI removal in different medical specialties?

Text-to-image models generate urban planning visuals, but they encode single perspectives on contested concepts like 'safety' and 'inclusivity' that vary across communities.

Method: The LIVS dataset contains 37,710 pairwise image comparisons across six criteria—Accessibility, Safety, Comfort, Invitingness, Inclusivity, and Diversity—collected from 30 community organizations over two years. Each comparison encodes community-specific preferences, revealing how different groups prioritize conflicting values. For example, one community might rank high-visibility lighting as 'safe' while another sees it as surveillance. The dataset structures these disagreements explicitly rather than averaging them away.

Caveats: Dataset covers urban planning contexts; transferability to other domains with value pluralism (healthcare, education) remains untested.

Reflections: How do you design interfaces that let users navigate between conflicting community values without defaulting to majority preferences? · Can models learn to predict which communities will disagree on specific criteria? · What happens when a single user belongs to multiple communities with conflicting values?

Expert tactile graphics designers spend hours customizing each chart due to fragmented workflows and complex guidelines. Non-experts can't start at all.

Method: Tactile Vega-Lite extends Vega-Lite with tactile-specific abstractions and synthesizes production-ready files for embossers and 3D printers. It provides smart defaults derived from formative interviews with expert designers—for example, automatically selecting appropriate texture patterns for categorical data or adjusting line weights for embossing constraints. The system outputs files in multiple formats (SVG for embossers, STL for 3D printing) without requiring designers to manually translate between tools.

Caveats: Smart defaults work for common chart types; highly specialized or novel visualizations still require expert customization.

Reflections: Can smart defaults adapt to individual reader preferences (e.g., texture sensitivity, spatial resolution)? · How do you version-control tactile chart specifications when guidelines evolve? · What's the learning curve for sighted designers with no tactile graphics experience?