📚 Research Digest

Finding: The AI placebo effect (believing AI is helping you changes performance) operates through two cognitive mechanisms: shifting expected performance levels AND reducing prior precision (confidence), with strong individual-level correlations (r=0.973) between expectation shifts and actual performance gains.

Why it matters: Directly relevant to HCI, behavioral science, and AI-human collaboration—demonstrates that *labeling* a tool as AI changes user cognition and behavior, with major implications for how AI-assisted tools should be designed and disclosed.

Method: Hierarchical Bayesian computational modeling with Leave-One-Out cross-validation to compare cognitive mechanism models; re-analysis of existing experimental data (N=75).

Jargon:
- *AI placebo effect*: Performance changes caused by belief in AI assistance, not the AI itself
- *Prior precision*: In Bayesian cognition, how confident/certain a person's expectations are (inverse of variance)
- *HDI (Highest Density Interval)*: Bayesian credible interval—range containing 95% of probable parameter values
- *elpd_diff*: Difference in expected log predictive density—measure of relative model fit in Bayesian comparison

Finding: Gamification mechanisms (via ClassDojo) in online phonics learning improved third-graders' learning outcomes, self-regulation (goal-setting, help-seeking, effort), and positive emotions (enjoyment, pride) compared to non-gamified online learning, with self-regulation strongly correlated with achievement (r=.73).

Why it matters: Directly relevant to gamification design, behavior change through engagement mechanisms, and goal-setting effects in learning contexts — all core teaching areas.

Method: Quasi-experimental design (n=52, ages 8-9) with experimental/control groups using Starfall + ClassDojo vs. Starfall alone; pre/post-test plus self-report measures.

Jargon: *ClassDojo* — a classroom management/gamification platform using points, badges, and feedback; *phonics* — letter-sound correspondence instruction; *self-regulation* — learners' ability to monitor and manage their own learning behaviors.

Finding: Introduces "Human-Centered AI Metacognition" (HC-AIM)—the capacity to monitor and regulate how much AI participates in one's own cognitive activities—and validates a psychometric scale measuring it across three studies.

Why it matters: Directly relevant to HCI and AI's impact on human cognition; the construct of preserving cognitive autonomy while collaborating with AI is critical for understanding skill acquisition, decision-making, and the design of human-AI systems.

Method: Three-study scale development and validation using exploratory factor analysis, confirmatory factor analysis, and convergent/criterion validity testing.

Jargon:
- *Metacognition*: Awareness and regulation of one's own thinking processes
- *Cognitive offloading*: Delegating mental tasks to external tools (here, AI)
- *AI literacy*: Understanding of how AI works and how to use it effectively
- *Cognitive autonomy*: Maintaining independent thinking rather than deferring to AI outputs

Finding: Humans and LLMs (ChatGPT, Gemini) differ in cross-modal associative matching — LLMs outperform humans at recovering intended word-image pairings, and LLMs are better at matching LLM-generated pairs than human-generated ones, suggesting partially distinct associative architectures despite surface similarities.

Why it matters: Directly relevant to understanding AI's creative capacities versus human creativity, with implications for HCI design tools, AI-assisted creative workflows, and how we teach creativity in an AI-augmented world.

Method: Experimental dyad generation + rematching paradigm comparing two LLMs against two human groups (generators vs. matchers), enabling clean comparison of cross-modal association across human and AI systems.

Jargon: *Cross-modal associations* = meaningful links formed between stimuli from different sensory/representational domains (here: words and images); *associative memory architecture* = the underlying structure by which a system (brain or model) stores and retrieves connected concepts.

Finding: People tend to "block" practice (repeating one task continuously before switching) in self-directed learning, and this behavior is better explained by a simple repetition bias than by existing models of intrinsic motivation based on novelty, competence, or learning progress.

Why it matters: Directly relevant to learning sciences, skill acquisition, and behavioral science—repetition bias as a cognitive tendency has implications for designing learning systems, gamification, and behavior change interventions that assume learners optimize for progress.

Method: Multi-task self-directed learning experiment comparing predictive power of epistemic motivation models vs. a repetition bias model.

Jargon: *Blocking* = practicing one task in a continuous block before switching; *interleaving* = alternating between tasks during practice; *epistemic factors* = knowledge/curiosity-driven motivators like novelty and perceived learning progress.

Finding: Otome game engagement (emotional and financial investment) correlates with identity development, social interaction, and romantic expectations among young adult Filipinas, but not with psychological well-being.

Why it matters: Touches on user engagement mechanisms, parasocial relationships, and in-game purchase motivation—relevant to gamification, engagement design, and potentially dark patterns in games.

Method: Quantitative correlational design (Spearman's rho, multiple linear regression) with 118 participants recruited via purposive/snowball sampling.

Jargon: *Otome games* — narrative romance games targeting women, featuring male love interests; *Parasocial interaction* — one-sided emotional bonds users form with media characters; *Uses and Gratifications Theory* — framework explaining why people actively seek specific media to satisfy needs.

Finding: Proposes a seven-component instructional framework for designing e-learning for life skills in primary schools, integrating cognitive load theory, dual-coding, multimedia learning, and scaffolding principles.

Why it matters: Touches on learning sciences and skill acquisition in digital environments, but stays at a purely theoretical/conceptual level with no empirical data or HCI design work.

Jargon: *Cognitive load theory*: idea that working memory has limited capacity, so instruction should minimize extraneous mental effort; *dual-coding*: learning is enhanced when verbal and visual information are presented together; *scaffolding*: temporary instructional support gradually removed as learners gain competence.

Finding: Teachers using generative AI for lesson design face challenges in adapting AI outputs to student needs and lack effective human-machine collaboration mechanisms; structured negotiation and evaluation cycles under teacher guidance improve outcomes.

Why it matters: Touches on AI-human collaboration in educational contexts and learning sciences, but is narrowly focused on high school English poetry teaching rather than HCI, creativity, or behavioral science broadly.

Method: Design-based research with semi-structured interviews of three teachers and two iterative practice rounds in a real classroom.

Jargon: *AI-TPACK* = framework extending Technological Pedagogical Content Knowledge to include AI literacy; *SAMR* = Substitution, Augmentation, Modification, Redefinition — a model for technology integration levels in education.

Finding: First-year college students' motivation (expectancies, utility value, cost) follows a recurring calibration pattern—declining mid-course then stabilizing—driven by course difficulty and performance feedback, with cumulative baseline drops across quarters signaling academic risk.

Why it matters: Tangentially relevant to goal-setting and motivation research, but focused narrowly on academic motivation trajectories rather than behavioral intervention, HCI, or applied behavioral science.

Method: Latent change score models across three quarters, tracking motivation three times per course in two courses per quarter (N=942).

Jargon: *Expectancy-value theory*: motivation framework where effort depends on expected success and perceived value/cost of tasks. *Latent change score models*: structural equation models that explicitly estimate within-person change over time.

Finding: Humans and LLMs both infer a four-dimensional structure of moral motivation from narratives, but LLMs underrepresent antisocial motivations and produce less varied moral judgments than humans.

Why it matters: Marginally relevant to AI-human alignment and behavioral science, but focuses narrowly on moral cognition rather than HCI, creativity, teams, or behavior change design.

Method: Compared 1,031 human participants against 23 frontier LLMs evaluating 344 moral narratives, using supervised ML to test dimensional models of motivation inference.