AI Moves From Tool to Infrastructure, Verification Lags Behind

AR guidance systems cause cognitive tunneling—users hyperfocus on virtual overlays and miss critical real-world hazards. In CPR scenarios, responders fixate on compression indicators while ignoring patient vomiting or environmental threats.

Method: Built a Magic Leap 2 app that tracks gaze patterns during AR-guided CPR to predict situational awareness failures. The system models when responders enter cognitive tunneling by measuring fixation duration on virtual compression guides versus real-world hazard zones. Uses eye-tracking data to classify awareness states in safety-critical moments.

Caveats: Study focused on CPR simulation; tunneling thresholds likely vary across domains and expertise levels.

Reflections: Can real-time gaze-based alerts break cognitive tunneling without disrupting task performance? · How do tunneling patterns differ between novices and experts in AR-guided procedures? · What overlay design patterns minimize tunneling risk while maintaining guidance effectiveness?

Designers rarely revisit early assumptions like personas or problem frames once they move to prototyping. Time constraints lock teams into initial choices, preventing iteration across interconnected design stages.

Method: StoryEnsemble uses forward-backward propagation to let designers edit any stage—persona, problem, solution, prototype—and automatically regenerate dependent artifacts. Change a persona's age from 25 to 65, and the system updates problem statements, feature ideas, and prototype screens to match. Built on LLM-generated design artifacts with bidirectional dependency tracking.

Caveats: Formative study with 15 participants; no longitudinal data on whether propagation actually improves design outcomes or just increases iteration volume.

Reflections: Does automatic propagation reduce designers' critical evaluation of AI-generated alternatives? · How do teams decide when to accept propagated changes versus manually overriding them? · Can this approach scale to enterprise design systems with hundreds of interdependent components?

BVI users need real-time visual assistance for navigation, object identification, and scene understanding. Existing solutions require human volunteers or lack interactive capabilities for dynamic environments.

Method: Tested ChatGPT's Advanced Voice with Video mode with eight BVI participants across real-world tasks. System failed at spatial precision—couldn't reliably distinguish 'left' from 'right' doors or provide accurate distance estimates. Succeeded at high-level scene description and text reading but struggled with fine-grained navigation cues critical for safe mobility.

Caveats: Exploratory study with eight participants using one LLM system; spatial reasoning may improve rapidly with model updates.

Reflections: Can prompt engineering or fine-tuning improve spatial reasoning for left-right distinctions? · How do BVI users calibrate trust when LLMs confidently provide incorrect spatial information? · What hybrid architectures combine LLM scene understanding with structured spatial reasoning?