AI Tutors Work Until Students Actually Use Them

Students use GenAI as a solution vending machine, not a learning tool, leading to worse exam performance. Telling them to 'use AI responsibly' doesn't teach them how to prompt for learning instead of answers.

Method: A semester-long RCT with 979 CS1 students tested four ICAP-based interventions of increasing cognitive engagement. All conditions significantly improved prompting skills, with gains increasing progressively from Condition 1 to Condition 4, validating ICAP's cognitive engagement hierarchy. Students with higher learning gains in immediate post-tests scored higher on final exams, though no direct between-group differences emerged.

Caveats: Learning gains predicted exam scores within groups, but between-group exam differences didn't materialize. Prompting skill ≠ automatic grade boost.

Reflections: What's the minimum intervention intensity needed to shift student behavior from solution-seeking to learning-oriented prompting? · Do prompting skills transfer across courses, or must each instructor re-teach them? · Why didn't between-group exam differences emerge despite improved prompting skills and within-group correlations?

Touch interactions require instrumented surfaces or computer vision, limiting where users can interact. Wearables could enable touch anywhere, but single-sensor approaches lack the fidelity to distinguish surface types, gestures, and tracking.

Method: TouchFusion fuses sEMG, bioimpedance, inertial, and optical sensing in a wristband to enable stateful touch detection on environmental and body surfaces, simple surface gestures, and trackpad-like tracking—without instrumenting surfaces. Validated on 100 participants to capture variance in wrist anatomies, skin conductivities, and behavioral patterns.

Caveats: Study validated detection and gestures but doesn't report accuracy rates for specific tasks. Real-world robustness unclear.

Reflections: What's the false positive rate for accidental touches during daily activities (typing, eating, gesturing)? · Can the system distinguish between intentional taps and incidental surface contact? · How does performance degrade with sweat, movement, or loose wristband fit?

LLMs absorb personal data during training and interactions, but users have no visibility into what models associate with their identity. Existing audits focus on memorization, not the associative inferences models make.

Method: LMP2 audits personal data associations across eight LLMs. For well-known individuals, models confidently generate multiple personal data categories. For everyday users, GPT-4o generates 11 features with 60% or more accuracy (gender, hair color, languages). 72% of participants sought control over model-generated associations with their name.

Caveats: Tested with EU residents. Accuracy and associations may differ for non-Western names or regions with less training data representation.

Reflections: Should data privacy rights extend to LLM-generated associations, even if they're inferred rather than memorized? · What counts as 'personal data' when a model generates plausible but incorrect associations? · Can users effectively contest or correct model associations without retraining?