📚 Research Digest

Finding: Pre-service teachers who produced higher-quality instructional designs with GenAI assistance showed a distinct evaluation-revision cycle ("planning → decision-making → generation → evaluation → revision → re-evaluation"), while lower-quality producers were stuck in fragmented, ineffective behavioral loops; cognitive load differences were large for mental demand (r = 0.647).

Why it matters: Directly relevant to AI's impact on creative/design work, human-AI collaboration patterns, and learning sciences—especially for a professor teaching HCI and AI's role in professional skill development.

Method: Lag sequential analysis compared behavioral transition sequences between high- and low-quality groups (median split on lesson plan scores), with multidimensional cognitive load measurement across 67 pre-service teachers.

Jargon: *Lag sequential analysis* = statistical method to identify which behavioral transitions occur more frequently than chance in sequential data; *cognitive load* = mental effort required by a task, typically split into intrinsic, extraneous, and germane components; *median split* = dividing participants into two groups at the midpoint of a score distribution.

Finding: People's different relational experiences with AI (therapeutic alliance, emotional attachment, social substitution) leave distinct—and sometimes undetectable—conversational traces in naturalistic ChatGPT interactions, with social substitution having the most identifiable linguistic fingerprint.

Why it matters: Directly relevant to AI's impact on human social behavior and collaboration, and raises important behavioral science questions about how people form parasocial/substitutive relationships with AI systems—critical for HCI design and ethics of persuasive/engaging AI.

Method: Combines self-reports from 94 adults with donated real ChatGPT conversations, blinded triple annotation of 500 sequences, and participant-grouped machine learning with pre-registered cross-validation—an unusually rigorous naturalistic design.

Jargon: *Digital Therapeutic Alliance (DTA)* — the perceived working relationship/bond between a user and an AI system, analogous to therapeutic alliance in psychotherapy. *Social Substitution* — using AI to replace human social relationships. *Ruptures* — breakdowns in the relational bond during conversation. *R²_CV* — cross-validated variance explained (out-of-sample predictive accuracy).

Finding: A French version of the Team Flow Monitor (TFM) is validated across two student team studies, confirming a four-facet structure (unity, joint progress, mutual trust, holistic focus) that predicts team performance and satisfaction but not individual performance.

Why it matters: Provides a validated instrument for researchers studying team dynamics and collective optimal experience, directly relevant to teaching about team collaboration and collective intelligence.

Method: Two-study validation using confirmatory factor analysis (CFA) with convergent/discriminant validity checks against work engagement, grit, and performance outcomes.

Jargon: *Team flow* — a shared optimal psychological state (analogous to individual "flow") where team members are fully immersed and performing collectively at their best; *CFI/RMSEA* — model fit indices for structural equation models; *second-order factor* — a latent variable that explains variance across multiple first-order subscales; *grit* — trait-level perseverance and passion for long-term goals, used here as a discriminant validity anchor.

Finding: A multimodal AI framework (MFIR-ME) evaluates classroom teaching quality by ranking acoustic, linguistic, visual, temporal, and behavioral features, achieving 88.6% accuracy and identifying linguistic coherence, student engagement, and teacher speech clarity as the most critical predictors.

Why it matters: Directly relevant to AI's impact on education and human-AI collaboration in learning environments, and the explainability dashboard + teacher annotation loop touches on meaningful human-AI teaming dynamics.

Method: Integrates SHAP, Random Forest, XGBoost, and LIME feature importance rankings via Borda Count voting consensus; uses Cross-Modal Attention Fusion on 236 recorded class sessions.

Jargon: *SHAP* = Shapley Additive Explanations (a method to explain individual model predictions); *LIME* = Local Interpretable Model-agnostic Explanations; *Borda Count* = a rank aggregation voting method; *MAE/RMSE* = mean absolute error / root mean square error (regression accuracy metrics).

Finding: Using Cox proportional hazards modeling on PISA data, the study builds a computational simulation to identify *when* students transition into academic burnout (amotivation), finding that longer weekly study hours accelerate burnout onset.

Why it matters: Marginally relevant to learning sciences and motivation, but the focus is narrowly methodological (survival analysis + simulation) with limited connection to HCI, behavior change design, or the professor's core teaching areas.

Method: Cox proportional hazards model applied to PISA data, with a ShinyApp for predicting burnout onset at individual/classroom/school levels.