📚 Research Digest

Finding: An LLM-based educational agent (DBagent) improved learning outcomes in a database course and produced distinctive cognitive engagement patterns—specifically a "Query-Evaluation-Query" verification loop—compared to traditional instructor-led interaction, with satisfaction mediating sustained engagement.

Why it matters: Directly relevant to AI's impact on learning, human-AI collaboration dynamics, and the design of AI-enriched educational environments, with implications for how autonomous agents differ from passive chatbots in pedagogical contexts.

Method: Quasi-experiment (n=313) combining lag sequential analysis (LSA) of interaction logs, SEM of survey data, and statistical comparison of learning outcomes across four authentic classes.

Jargon: *Lag sequential analysis (LSA)* — a method for identifying statistically significant behavioral sequences in time-ordered interaction data; *SEM (Structural Equation Modeling)* — statistical technique modeling causal relationships among latent variables; *cognitive engagement levels* — categorization of learner thinking depth (e.g., surface vs. deep processing).

Finding: Turn-duration burstiness (irregular temporal distribution of speech) in multidisciplinary oncology team meetings correlates with case complexity and processing time, suggesting it can serve as a quantitative marker of collective cognitive load.

Why it matters: Directly relevant to team dynamics and collective intelligence — offers a novel, measurable signal for real-time monitoring of team cognitive demand, with implications for how teams process complex decisions under naturalistic constraints.

Method: Burstiness parameter metrics (B and Aₙ(r)) applied to speech turn data, with random intercept models and growth curve modeling to link temporal communication structure to case complexity and processing time.

Jargon: *Burstiness* — a statistical measure of how irregularly events (here, speech turns) cluster over time vs. a random (Poisson) distribution; *MDTMs* — multidisciplinary team meetings; *growth curve modeling* — a longitudinal statistical approach tracking change over time/sequence.

Finding: Higher stakes in AI-assisted decision-making prolong deliberation but don't uniformly improve advice integration or accuracy—instead, effects depend on cognitive ability, with higher-ability people improving and lower-ability people performing worse under high stakes.

Why it matters: Directly relevant to behavioral science (motivation ≠ performance), AI-human collaboration, and the design of incentive structures in decision-support systems—important for teaching about cognitive biases, advice-taking, and AI integration in work.

Method: Two preregistered experiments (N=99) using Bayesian optimal weighting benchmarks to evaluate advice integration quality across varied stake conditions.

Jargon: *Advice integration* = how people combine their own judgment with external recommendations; *Bayesian benchmark* = mathematically optimal advice weighting used as performance standard.

Finding: People in Reddit mental health communities use AI (especially ChatGPT) across all four functional categories of social support (emotional, informational, appraisal, instrumental), but post-launch discourse shifted toward more negative, anxious, and clinically salient themes, with therapy-substitute and existential-threat framings driving the most engagement.

Why it matters: Directly relevant to AI's impact on human behavior and social interaction, behavioral science around trust in AI agents, and ethical design of persuasive/supportive technology—especially the "Computers Are Social Actors" framing and vulnerable user exploitation risks.

Method: Structural topic modeling on 10,042 Reddit posts + LIWC affective language scoring + negative binomial regression for engagement estimation, with a matched non-AI baseline for comparison.

Jargon: *CASA (Computers Are Social Actors)*: the theory that humans apply social rules to computers as if they were people. *LIWC*: Linguistic Inquiry and Word Count, a tool measuring emotional tone in text. *Structural topic modeling*: a method that identifies latent themes in text while modeling how covariates affect topic prevalence.

Finding: AI tools create a "dual digital vulnerability" where socioeconomically disadvantaged children face greater cognitive/relational risks from AI while having less access to its educational benefits, potentially harming theory of mind and self-regulation development.

Why it matters: Directly relevant to AI's societal impact, human-AI relationship dynamics, and the behavioral/developmental risks of conversational AI designed for engagement—connects to persuasive technology ethics and AI's effects on human collaboration and cognition.

Method: Synthesizes empirical data from ARCOM, ANCT-CRÉDOC, and OECD alongside social media research literature to build a public health argument about AI and child development.

Jargon: *Theory of mind* = ability to attribute mental states to others; *symbolic functioning* = capacity to use symbols/language for abstract thought; *ARCOM/ANCT-CRÉDOC* = French regulatory and research bodies on digital media.

Finding: A smartphone app (HippoCamera) that guides users to record and review memory cues produces lasting episodic memory improvements in older adults, persisting 3.5–6 years after the intervention, with benefits extending to well-being and daily lifestyle.

Why it matters: Directly relevant to behavior change design and persuasive technology—demonstrates long-term efficacy of a neuroscience-guided mobile intervention using established mnemonic strategies, with implications for HCI and habit formation research.

Method: Mixed-methods longitudinal follow-up (n=25) combining subjective re-experiencing ratings with semi-structured interviews, comparing reviewed vs. recorded-only vs. calendar-cued events.

Jargon: *Episodic recollection* = memory for specific personal past events; *hippocampal activity* = neural activity in the brain region central to memory formation.

Finding: Calibrates cybersickness severity scales (SSQ, FMS) against actual user dropout behavior in VR, creating behaviorally grounded risk bands that predict when users will quit a VR experience.

Why it matters: Tangentially relevant to VR usability/HCI and user engagement/dropout mechanisms, but focused narrowly on motion sickness measurement rather than interaction design, behavior change, or the professor's core topics.

Method: Logistic regression + survival analysis linking subjective sickness ratings to dropout probability across two studies (N=183, N=304).

Finding: A webcam-based assistive HCI system uses facial recognition, eye-head tracking, and blink gesture recognition to enable computer control for people with physical limitations.

Why it matters: Marginally relevant to HCI and accessibility design, but focused narrowly on assistive technology engineering rather than the behavioral, cognitive, or design dimensions the professor teaches.