AI Assistance Breaks Down When Users Can't Verify Outputs

Algorithmic matching systems in healthcare and social services don't improve human decisions. Humans using AI often perform worse than either the human or algorithm alone.

Method: The researchers built confidence-aware interfaces that show when the algorithm is uncertain. The system uses a complementarity score—measuring whether human-AI collaboration beats solo performance—and surfaces cases where human judgment adds value. In their healthcare matching experiments, the interface flags low-confidence predictions and lets humans override with domain knowledge the algorithm lacks.

Caveats: Requires ground-truth data to calibrate confidence scores. Won't work in domains where you can't measure prediction certainty.

Reflections: How do you design confidence displays that don't overwhelm users with uncertainty information? · What's the optimal threshold for routing decisions to humans vs. algorithms? · Can complementarity scores generalize across different matching domains?

Blind users can't accurately measure bar heights or line positions in tactile graphics. Visual design guidelines fail when translated to touch.

Method: Through controlled experiments with 13 blind participants, the study found that tactile bar chart reading accuracy drops to 60% compared to 90%+ for visual charts. Users developed compensatory strategies: anchoring fingers at axis endpoints, using both hands to span distances, and verbalizing measurements aloud to maintain spatial memory. The key failure: visual encodings assume simultaneous perception of multiple elements, but tactile perception is sequential.

Caveats: Study used simplified bar charts. Complex multi-series visualizations or scatterplots remain untested.

Reflections: How do you encode relative comparisons (e.g., trend direction) in tactile form? · Can haptic feedback augment static tactile graphics for measurement tasks? · What's the cognitive load difference between visual and tactile chart reading?

Ray-casting in dense MR environments hits the wrong object 40% of the time. Occluded or clustered targets are nearly impossible to select accurately.

Method: The system maps finger curvature to real-time Bezier curves for expressive selection paths, then projects the four nearest candidates onto the user's forearm using proximity-based spatial mapping. Users bend their finger to trace a curve through 3D space, then tap their forearm to pick the intended target. In user studies, this reduced selection errors by 68% compared to standard ray-casting in cluttered scenes.

Caveats: Requires forearm tracking. Won't work with long sleeves or for users with limited arm mobility.

Reflections: How does forearm mapping scale beyond four candidates? · Can this technique generalize to other body surfaces like the palm? · What's the learning curve for non-technical users?