📚 Research Digest

Finding: A verification-oriented AI instructional framework (GAIIM) where students solve problems independently then compare, detect errors in, and verify AI-generated solutions significantly outperformed conventional instruction on algebraic problem-solving, error detection, and error resolution over 12 weeks.

Why it matters: Demonstrates a concrete pedagogical model for AI integration that promotes critical thinking rather than answer-seeking—directly relevant to teaching AI's impact on learning and skill acquisition, and to designing effective human-AI collaboration frameworks.

Method: Explanatory sequential mixed-methods: randomized pretest-posttest control group (n=86) followed by semi-structured interviews (n=10) to elaborate quantitative findings.

Jargon: *Verification-oriented*: students evaluate and critique AI outputs rather than use AI to generate answers. *GAIIM*: the specific instructional sequence of solve → compare → detect errors → resolve → reflect.

Finding: A multimodal AI framework (MFIR-ME) evaluates classroom teaching quality by ranking acoustic, linguistic, visual, temporal, and behavioral features, achieving 88.6% accuracy in predicting teaching quality scores, with linguistic coherence, student engagement, and teacher speech clarity identified as top predictors.

Why it matters: Directly relevant to AI's impact on education and human-AI collaboration in learning environments, plus the explainability dashboard and teacher feedback loop touch on HCI design principles for AI-assisted tools.

Method: Uses SHAP, Random Forest, XGBoost, and LIME feature importance rankings aggregated via Borda Count voting, combined with a Cross-Modal Attention Fusion module across 236 class sessions (~11,000 minutes of video).

Jargon: *SHAP* (SHapley Additive exPlanations) = method to explain individual feature contributions to model predictions; *LIME* = Local Interpretable Model-agnostic Explanations, another local explainability method; *Borda Count* = a rank aggregation voting mechanism; *MAE/RMSE* = Mean Absolute Error / Root Mean Square Error, regression accuracy metrics.

Finding: A web-based debiasing intervention (teaching Bayesian reasoning and debiasing strategies) significantly reduced decision-making errors and improved diagnostic accuracy among psychologists, with stronger effects for probabilistic than categorical diagnostic tasks.

Why it matters: Directly relevant to behavioral science (cognitive bias debiasing, decision-making errors) and behavior change design (web-based intervention effectiveness), and touches on learning sciences (skill acquisition via online training).

Method: Randomized controlled experiment (N=535 Brazilian psychologists) with vignette-based outcome measures; active control condition used.

Jargon: *Debiasing* = interventions designed to reduce systematic cognitive errors; *Bayesian probability estimates* = updating diagnostic likelihood based on base rates and evidence.

Finding: Human-AI collaboration events (novelty, disruption, criticality) affect students' information security behavior through dual pathways: security vigilance (positive) and cognitive overload (negative).

Why it matters: Relevant to AI's impact on learning environments and the behavioral science of how cognitive load and vigilance mediate technology-induced behavior change.

Method: Three-wave time-lagged survey (n=351) with mediation analysis through dual pathways.

Jargon: *Event system theory* — framework treating events (rather than stable states) as drivers of organizational/behavioral change; *event strength* — degree to which an event's novelty, disruption, and criticality compel attention and response.

Finding: Fixation count is more sensitive than blink or pupil measures for distinguishing attention types (global, selective, sustained) in low-demand visual tasks during HCI.

Why it matters: Relevant to attention-aware HCI systems design, but the study is a small pilot (n=10) with marginal effects and limited practical takeaways for teaching or research in UX, learning, or behavior change.

Method: Repeated-measures eye tracking with static image tasks; very small sample limits generalizability.

Finding: GenAI is reshaping entrepreneurship by supporting idea generation, business model development, and decision-making, but introduces risks like overdependence, misinformation, and unequal digital access; the paper proposes "AI entrepreneurial capability" as a key competency framework.

Why it matters: The human-AI collaboration angle and the concept of AI literacy/capability are tangentially relevant to courses on AI's impact on work and creativity, but the entrepreneurship focus limits direct applicability.

Jargon: *AI entrepreneurial capability* — a proposed construct combining AI literacy, prompt engineering, output evaluation, and responsible AI use in business contexts; *SMEs* — small and medium enterprises.

Finding: A hybrid CNN–Random Forest model with residual connections and channel attention achieves real-time facial emotion recognition across seven emotion classes.

Why it matters: FER systems are relevant to HCI and affective computing applications, but this paper is primarily a technical deep learning architecture contribution with minimal engagement with HCI design, user experience, behavior, or learning contexts.

Jargon: *Residual connections*: shortcuts in neural networks that help gradients flow during training; *Channel attention*: a mechanism that weights feature map channels by importance; *Affective computing*: computing that recognizes or simulates human emotions.

Finding: Traditional sports games significantly increased fourth-grade students' physical education (PJOK) learning motivation, with mean scores rising from 67.11 to 82.48 post-intervention.

Why it matters: Marginally relevant to games and motivation in learning contexts, but focused narrowly on physical education in Indonesian elementary schools with limited applicability to HCI, creativity, or behavioral science research.

Method: Pre-experimental one-group pretest-posttest design with Likert-scale questionnaire and paired sample t-test (n=27).

Finding: The SAVI (Somatic, Auditory, Visualization, Intellectually) learning model significantly improved elementary students' physical education outcomes, with pretest scores rising from 50.66 to 78.50 post-intervention.

Why it matters: Tangentially relevant to learning sciences and multimodal/multisensory instructional design, but narrowly focused on physical education in Indonesian elementary schools with a weak pre-experimental design.

Method: One-group pretest-posttest design (no control group), limiting causal inference.