Users Can't Report Their Own Preferences Accurately

Users claim urgency drives their need for AI writing help, but their actual behavior tells a different story—creating a design paradox for proactive assistants.

Method: A factorial vignette study with 750 pairwise comparisons reveals compositional effort dominates user decisions with a correlation of ρ=0.597, while urgency—the factor users self-report as most important—shows zero predictive power (ρ≈0). The perception-behavior gap is stark: users rank urgency first in surveys despite it being the weakest behavioral driver. This means self-reported preferences are worse than useless for training proactive AI.

Caveats: Study used vignettes, not live writing sessions. Real-time compositional effort detection remains an unsolved instrumentation problem.

Reflections: Can real-time compositional effort be reliably detected from keystroke dynamics alone? · Do perception-behavior gaps persist after users experience multiple AI interventions? · What other domains exhibit similar misalignment between stated and revealed preferences for AI assistance?

XR productivity apps fail not because of motion sickness or latency, but because text clarity is fundamentally broken—users get sick trying to read.

Method: Controlled experiments isolating effective resolution (the clarity after display-optics-rendering pipeline) show it's the primary driver of simulator sickness during text-heavy tasks. The study systematically degrades resolution while holding latency and motion constant, proving that sub-optimal text rendering causes physiological discomfort independent of vestibular conflicts. Reading performance degrades proportionally with resolution loss, creating a direct link between visual fidelity and usability.

Caveats: Study focused on static text reading tasks. Dynamic content with motion may introduce additional sickness vectors beyond resolution.

Reflections: What is the minimum effective resolution threshold for sustained XR productivity work? · Do adaptive text rendering techniques that boost clarity reduce sickness without performance penalties? · Can perceptual training reduce resolution-induced sickness over extended XR exposure?

LLM-based smart home control is unreliable and expensive—every command requires full model inference, creating latency and unpredictability that breaks the user experience.

Method: The system distills LLM reasoning into a lightweight finite-state machine that handles routine commands locally, reserving full LLM inference only for ambiguous or novel requests. It learns user-specific device preferences (e.g., 'dim the lights' maps to 30% for one user, 50% for another) and compiles them into deterministic rules. This hybrid architecture cuts inference latency by 73% and cost by 68% while maintaining natural language flexibility for edge cases.

Caveats: Requires initial training period to learn user preferences. Cold-start users get generic mappings until personalization kicks in.

Reflections: How many interactions are needed before personalized rules outperform generic LLM responses? · Can the finite-state machine gracefully degrade when user preferences shift over time? · What percentage of smart home commands are truly novel versus routine across different user populations?