📚 Research Digest

Finding: Self-regulated learning quality determines who benefits from LLMs—users who steer, monitor, and repair LLM outputs build genuine competence, while those who offload synthesis accumulate "verification debt" and develop false mastery (better assisted performance without actual learning transfer).

Why it matters: Directly addresses AI's impact on learning, skill acquisition, and human-AI collaboration, with implications for HCI design (cognitive workbenches), behavioral science (confidence calibration, cognitive offloading), and the creativity/expertise literature.

Method: Theoretical framework mapping self-regulated learning loops onto observable interaction traces (constraint density, monitoring calibration, repair depth) as measurable process markers.

Jargon:
- *Synthesis trap*: offloading comparison/integration to LLM, producing fluent output without understanding
- *Verification debt*: accumulated unchecked LLM outputs that undermine error detection capacity
- *Cognitive workbenches*: interface designs that externalize standards and preserve user agency over verification
- *Regulatory support*: interface features that scaffold self-monitoring and goal-setting behaviors

Finding: A Meta Quest 3-based mixed reality serious game for diagnostic classification in higher education achieved learning outcomes comparable to expert-guided field training, outperforming classroom instruction and self-directed study, with demonstrated real-world transfer.

Why it matters: Directly relevant to serious games, gamification, learning sciences, and skill acquisition—provides both design recommendations and a methodological framework for studying MR game-based learning that could inform teaching on games and HCI.

Method: Two complementary studies with undergraduates comparing instructional approaches, using triangulated evidence (objective metrics, questionnaires, interviews) within a mixed-method design.

Jargon: *Mixed Reality (MR)*: technology blending physical and digital environments (beyond VR/AR); *Transfer*: application of learned skills to real-world tasks outside the training context; *Gamification*: applying game mechanics to non-game contexts to drive engagement.

Finding: AI agents can pass conventional data quality checks in online research, creating systematic "contamination" through three pathways (full delegation, partial mediation, spillover) that distort survey and crowdwork data while appearing human-generated.

Why it matters: Directly challenges the validity of behavioral science research conducted online—any professor teaching replication, behavioral economics, or research methods needs to understand how AI contamination threatens the evidential foundation of the field.

Method: Synthesis of emerging evidence with a proposed taxonomy of contamination pathways and vulnerability zones, plus a minimum reporting checklist for online studies.

Jargon: *Contamination pathways*: ways AI shapes responses (full = AI answers entirely; partial = AI assists; spillover = prior AI use affects later behavior); *adversarial robustness*: resistance to deliberate attempts to defeat detection systems.

Finding: A preregistered RCT (N=344) found that both a one-hour cognitive apprenticeship-based lateral reading training and a one-page self-guided version improved credibility judgments, analytic thinking, and persuasion knowledge compared to controls.

Why it matters: Directly relevant to behavioral science (persuasion knowledge, cognitive biases), behavior change design (intervention design), and learning sciences (skill acquisition for critical evaluation), with implications for AI literacy and misinformation resistance.

Method: Preregistered RCT with Bayesian regression analyses comparing two active intervention conditions against a control group.

Jargon: *Lateral reading* = checking a source's credibility by opening new browser tabs to search what others say about it, rather than reading the source itself. *Persuasion knowledge* = awareness of tactics used to manipulate or mislead. *Cognitive apprenticeship* = instructional method where an expert models thinking processes for learners.

Finding: All five state-of-the-art LLMs systematically overrate their own moral reasoning relative to human baselines across every wisdom dimension, an artifact linked to RLHF training that produces inflated, variance-suppressed responses insensitive to scenario context.

Why it matters: Directly relevant to AI impact on high-stakes human roles and raises critical questions about deploying LLMs in collaborative, therapeutic, or decision-support contexts where genuine moral reasoning matters.

Method: Dual-layer pipeline combining lexical-emotional profiling with a behaviorally anchored 5-dimension wisdom rubric, validated against human expert raters (ICC = .764) before benchmarking LLMs; includes a self-serving-bias check using LLMs as raters of their own outputs.

Jargon: *ICC (Intraclass Correlation Coefficient)*: inter-rater reliability measure; *RLHF (Reinforcement Learning from Human Feedback)*: training method that fine-tunes LLMs using human preference ratings, hypothesized here to cause systematic response inflation; *alignment artifact*: a systematic distortion introduced by the training process rather than reflecting genuine capability; *cross-vignette variance paradox*: LLMs show artificially low variance in scores across different moral scenarios despite humans showing meaningful differentiation.

Finding: A gamified digital physical activity intervention (KiplinTM) for obese diabetic patients showed no significant advantage over usual care in autonomous motivation or weight stigma reduction, and proposed theory-based psychological mechanisms (self-determination theory, social identity) were not supported as mediators.

Method: Randomized controlled trial (N=50, 3-month intervention, 6-month follow-up) using mixed-effects models and partial least squares structural equation modeling (PLS-SEM) to test mediation.

Why it matters: Directly relevant to gamification design and behavior change research—the null findings on motivational mechanisms challenge assumptions about *why* gamified interventions work and have implications for persuasive technology and habit formation design.

Jargon: *Autonomous motivation*: doing an activity because it's personally meaningful (vs. externally pressured); *Self-determination theory (SDT)*: framework linking need satisfaction (autonomy, competence, relatedness) to motivation; *PLS-SEM*: structural equation modeling approach suited for small samples and exploratory mediation testing.

Finding: Cognitive diagrams (like Baddeley's working memory model) embed systematic biases—representing abstract/distributed processes as small, discrete, bounded objects—which constitutes a structural category error that shapes theory construction and experimental design.

Why it matters: Directly relevant to visual thinking, sketching, and design (how diagrams encode theoretical commitments), learning sciences (how textbook representations shape understanding), and cognitive bias (how representational conventions mislead reasoning).

Method: CODA (Conditions, Operations, Distortions, Actions) qualitative textbook analysis comparing cognitive and instrumentation diagrams.

Jargon: *Image schemas*: recurring spatial/bodily patterns (like CONTAINER or LARGE/SMALL) that structure abstract thought per embodied cognition theory. *Pen-model bias*: the tendency to draw abstract processes as small, bounded, discrete objects due to 2D diagrammatic constraints. *Countable-object ontology*: treating uncountable/distributed processes as if they were discrete, enumerable things.

Finding: Artists can deliberately design artworks to be memorable or forgettable, but AI (deep neural network ResMem) outperforms both artists and laypeople in predicting which artworks will be remembered, using psychologically interpretable visual features.

Why it matters: Touches on creativity and design (visual memorability as a designable property), AI's superiority over human intuition in perceptual prediction, and has mild relevance to HCI/UX design where memorability matters—but the core contribution is more perceptual psychology than any of the professor's primary teaching domains.

Method: Nationwide art contest generating intentionally memorable/forgettable artworks, tested via online and in-gallery recognition tasks, with clustering/regression analyses and AI model comparison.

Jargon: *ResMem*: a deep neural network trained to predict image memorability scores. *Memorability*: the probability that an image will be recognized after a delay, shown to be consistent across viewers.

Finding: In a large-scale mobile eye-tracking study, shoppers use perceptually accessible attributes (especially brand) to guide visual search in multi-option decisions, with attribute-based search increasing as choice set size grows—consistent with adaptive Bayesian updating.

Why it matters: Directly relevant to choice architecture and behavioral economics (how consumers navigate complex choice environments), with implications for nudge design, dark patterns, and decision-making under cognitive load.

Method: Mobile eye-tracking in a controlled mock grocery store (n > 200) combined with a Bayesian updating model of efficient search.

Jargon: *Attribute-guided search* — using a salient feature (e.g., brand color/logo) as a filter before evaluating other product attributes; *Bayesian updating* — sequentially revising probability estimates as new information is encountered.

Finding: LLMs used as psychological research tools (simulators, cognitive models) produce "measurement phantoms"—artifacts that mimic genuine psychological phenomena—and the authors propose a six-stage validity workflow to distinguish real computational phenomena from measurement noise.

Why it matters: Relevant to AI's impact on behavioral science research methodology, particularly for researchers using LLMs to simulate human participants or measure psychological constructs—a growing practice that affects the credibility of behavioral and HCI research.

Method: Proposes a dual-validity framework integrating psychometrics with causal inference, illustrated through an "LLM selfhood" case study.

Jargon: *Measurement phantoms*: statistical artifacts in LLM outputs that mimic psychological phenomena but are caused by trivial prompt variations; *construct validation*: verifying that a measurement instrument actually captures the intended psychological construct.

Finding: Reanalysis showing that cross-lagged panel model results linking internet gaming disorder and sleep problems may be statistical artifacts rather than genuine causal relationships, using multiverse methodology across competing models.

Why it matters: Directly relevant to behavioral science and replication/methodology concerns—demonstrates how analytic choices can produce conflicting conclusions about gaming disorder, which matters for teaching statistical reasoning and causal inference in behavioral research; tangentially touches games and behavior change.

Method: Simulated data reproducing the original correlation structure, then fitted multiple competing models (cross-lagged panel, spurious association models) as a multiverse analysis.

Jargon: *Cross-lagged panel model*: longitudinal statistical model estimating whether variable A at time 1 predicts variable B at time 2 (and vice versa). *Multiverse methodology*: running all defensible analytic alternatives to assess robustness of findings.

Finding: A structured self-regulated learning (SRL) intervention (goal-setting, self-monitoring, reflection) added to AI-assisted vocabulary learning significantly improved both vocabulary achievement and enjoyment compared to unguided AI use, with stronger effects for low-enjoyment students.

Why it matters: Touches on goal-setting scaffolding, AI's impact on learning, and behavior change design (structured routines vs. autonomous tool use), but the domain (ESP vocabulary instruction) is narrow and applied, limiting generalizability to the professor's core interests.

Method: Quasi-experimental design with 266 students across pre-existing classes; moderation analysis by baseline enjoyment level.

Jargon: *SRL (Self-Regulated Learning)* — metacognitive process of setting goals, monitoring progress, and reflecting on performance; *ESP (English for Specific Purposes)* — English instruction tailored to professional/technical domains.

Finding: Bibliometric analysis of 263 papers (1980-2025) traces how managerial decision-making theory evolved through technological eras—from bounded rationality through data-driven and algorithmic configurations to agentic AI systems.

Why it matters: The shift toward algorithmic bias, explainability, and autonomous decision governance touches on AI's impact on work and behavioral decision-making, but the managerial/organizational focus is somewhat peripheral to the professor's core interests.

Method: Bibliometric-conceptual design combining VOSviewer keyword co-occurrence mapping with qualitative historical interpretation—a hybrid quantitative/interpretive synthesis approach.

Jargon: *Bounded rationality*: Herbert Simon's concept that human decision-making is limited by cognitive capacity and information availability. *Agentic intelligence*: AI systems capable of autonomous goal-directed action. *Thematic evolution analysis*: tracking how research themes shift over time using bibliometric data.

Finding: Survey study finding that users generally value both explainability and controllability in AI systems, but show high variance in opinions, suggesting individual differences (technical knowledge, context, experience) shape perceptions of trustworthy human-AI collaboration.

Why it matters: Touches on AI impact on human experience and HCI design principles, but the findings are thin and generic — no novel behavioral or design insights that would advance teaching in HCI, behavioral science, or AI-human collaboration.

Method: Online survey measuring 12 aspects of explainability and controllability across users; built on a prior pilot study.

Finding: Psychology's WEIRD sampling problem stems from "disciplinary selectionism" — the privileging of Euro-American institutional and philosophical traditions that has systematically excluded non-Western theories of mind and constrained the field's explanatory frameworks.

Why it matters: Tangentially relevant as a meta-critique of behavioral science's foundational assumptions, which could inform how the professor contextualizes generalizability claims in behavioral science and decision-making research.

Jargon: *WEIRD* = Western, Educated, Industrialized, Rich, Democratic (describing non-representative research samples); *disciplinary selectionism* = the process by which privileging certain intellectual traditions narrows the theories and methods a field develops.

Finding: EEG brain responses to AI-generated faces can be used to train classifiers that reveal implicit social biases (e.g., masculine/senior leadership stereotypes), with BrainVote-generated images winning mock elections at 2.67x the rate of controls.

Why it matters: Tangentially relevant to HCI (brain-computer interfaces, AI-human interaction) and behavioral science (implicit bias measurement), but the core focus is neuroscience/social psychology methodology rather than the professor's primary teaching areas.

Method: Novel BCI pipeline combining EEG-based neural classifiers with generative AI image recalibration to externalize implicit preferences — technically impressive but niche.

Jargon: *EEG evoked potentials* = electrical brain signals recorded in response to stimuli; *Implicit Association Test (IAT)* = reaction-time measure of unconscious bias; *generative neural network* = AI model that creates synthetic images.

Finding: Positive incentives (rewards) drive faster responses and approach motivation, while negative incentives (penalties) drive slower, more accurate responses and avoidance/threat states; affective states mediate effort allocation differently depending on incentive type.

Why it matters: Has marginal relevance to motivation design, gamification reward structures, and behavioral science around incentives and loss aversion, but the focus is narrow psychophysiology rather than applied behavioral design.

Method: Two-study experimental design with psychophysiological measurement during cognitively demanding tasks under reward/penalty conditions, with stress state manipulation in Study 2.

Jargon: *Challenge/threat states*: cardiovascular response patterns reflecting positive vs. negative appraisal of demands; *sympathetic arousal*: autonomic nervous system activation associated with stress responses.

Finding: A microlearning-based digital storytelling model (MICRO-DST) significantly improved English speaking skills in Thai EFL students, with a large effect size (Cohen's d = 1.51) and ~48% proficiency gain.

Why it matters: Tangentially relevant to learning sciences and skill acquisition, and microlearning/digital storytelling touch on instructional design, but the domain (EFL language instruction) is far from the professor's core interests.

Method: One-group pretest–posttest quasi-experimental design with n=35; no control group limits causal inference.