Systems Optimize Actions While Missing Intent

Drivers must decide when to engage driving automation and when to take back control—a judgment that imposes steep cognitive load and creates safety risks from both over-reliance and delayed intervention.

Method: Synchronized 136.6 hours of naturalistic driving across 127 drivers, capturing front-view video, in-cabin video, CAN bus signals, radar, and GPS context around each control transition. Visual input alone failed: front-view video missed driver state, in-cabin video missed road context. Adding CAN and route signals substantially improved prediction over video-only baselines. Takeover events developed gradually and benefited from longer horizons; handover events depended on immediate cues—an asymmetry with direct HMI implications.

Caveats: Benchmark tasks defined but baseline performance leaves room for improvement. Real-time deployment latency not tested.

Reflections: What prediction horizon minimizes false alarms while maintaining safety margins for takeover events? · Do transition patterns generalize across vehicle types and automation levels (L2 vs L3)? · Can driver-specific models trained on individual transition history outperform population-level baselines?

AI coding assistants optimize for developer speed, not energy impact. Frontend code replicated across millions of page views consumes significant energy, yet existing sustainability guidelines see limited practitioner adoption.

Method: An IDE-integrated assistant that analyzes AI-generated frontend code, estimates energy footprint, and proposes optimizations. Benchmarked on 500 websites and tested with 20 developers. Reduced per-website energy by 13-16% on average while maintaining developer productivity. Increased developers' awareness of energy use through actionable, in-context feedback during the coding workflow.

Caveats: Tested only on frontend code. Backend, mobile, and embedded systems may have different energy profiles and optimization opportunities.

Reflections: Do energy savings persist as developers internalize optimization patterns, or does awareness decay without continuous feedback? · How do energy-optimized implementations affect user-perceived performance and accessibility? · Can similar approaches extend to backend code where energy attribution is more complex?

Lead marketing platforms collect and sell personal data submitted via web forms, yet the ecosystem's privacy and spam risks remain largely unmeasured despite handling sensitive health and financial information.

Method: Instrumented over 100 health-related lead-generation websites and monitored 200 controlled phone numbers and emails. Observed sharing of highly personal health information to more than 70 distinct third parties. Purchased leads from three major platforms, uncovering deceptive brokerage where consumer data was sold to unvetted buyers and often augmented or fabricated with attributes like health status. Received over 8,000 telemarketing calls, 600 texts, and 200 emails—calls often began within seconds of form submission. Phone-based opt-outs helped most but none completely stopped marketing.

Caveats: Focused on health-related lead generation. Other verticals (insurance, education) may have different broker practices and regulatory compliance.

Reflections: What technical or regulatory interventions could enforce data minimization and buyer vetting in lead platforms? · How do consumers perceive the trade-off between quote convenience and downstream marketing harassment? · Can browser-level interventions detect and warn users before submitting data to lead-generation forms?