📚 Research Digest

Finding: Labeling identical creative exemplars as human- vs. AI-generated significantly changes how dyads engage with them—human attribution increased idea uniqueness and interbrain synchronization, while AI attribution increased pathway switching but neither effect persisted after exemplar withdrawal.

Why it matters: Directly relevant to AI's impact on creativity education, team collaboration, and the behavioral science of attribution effects—with strong implications for how educators and designers should frame AI-generated instructional content.

Method: fNIRS hyperscanning (dual-brain neuroimaging) with 113 dyads doing creative ideation tasks—an unusually rigorous method for studying collaborative creativity that captures real-time neural synchrony between partners.

Jargon: *fNIRS hyperscanning* = simultaneous near-infrared brain scanning of two people to measure interbrain synchronization; *IBS (interbrain synchronization)* = degree to which two brains show correlated neural activity during collaboration; *exemplar-category overlap* = how much generated ideas fall within the same category as shown examples; *nearest-exemplar semantic distance* = how conceptually far ideas are from the provided examples.

Finding: Creative engagement (especially creative writing) increases subsequent financial risk-taking behavior, but neither self-efficacy nor affective dominance explains this relationship, leaving the mechanism unclear.

Why it matters: Directly connects creativity research to decision-making and behavioral science, relevant for understanding downstream effects of creative activities on risk cognition—useful for teaching about creativity, behavioral science, and design of creative interventions.

Method: Three-study design combining cross-sectional survey, experimental creative drawing vs. coloring task (binary lottery), and creative writing vs. standard typing task (investment-choice task).

Jargon: *Affective dominance* = the sense of feeling in control or powerful emotionally; *risk-taking propensity* = general tendency to accept risk across domains.

Finding: The DRC (divergence-reflection-convergence) framework reduces cognitive bias in LLMs at inference time without retraining, by generating competing viewpoints, reflecting on them, and converging on a less biased output.

Why it matters: Directly relevant to the user's interest in AI's societal impact and cognitive biases—LLMs embodying and amplifying human cognitive distortions is a topic the user has already flagged as highly relevant.

Method: Tests five LLMs across four cognitive bias dimensions using asymmetric kappa (AK) and total variation distance (TVD) as evaluation metrics.

Jargon: *Inference-time strategy* = bias correction applied during model use, not during training. *Asymmetric kappa (AK)* = modified inter-rater agreement metric for skewed/subjective judgments. *TVD (total variation distance)* = measure of how different two probability distributions are, used here to quantify output inconsistency.

Finding: Cybercriminals exploit cognitive biases, trust mechanisms, impulsivity, and emotional responses to victimize individuals; psychological traits like trust propensity, risk perception, and digital literacy predict vulnerability.

Why it matters: Directly relevant to behavioral science, cognitive biases, and decision-making — and touches on persuasive technology's dark side (social engineering, dark patterns, manipulation) which maps to the behavior change design interest area.

Method: Theoretical/review synthesis drawing on victimology, cyberpsychology, and criminology frameworks.

Finding: Interactive video lessons with embedded questions, immediate feedback, and gamified elements improved both motivation and English proficiency among university students in a workplace English course.

Why it matters: Touches on gamification, engagement mechanics, and motivation constructs (goal orientation, self-efficacy, task value) relevant to behavior change design and learning sciences.

Method: Mixed-methods, pre/post design with no control group — weak causal inference.

Jargon: *iSLCollective* — a platform for customizing video lessons with interactive overlays; *goal orientation* — whether learners are motivated by mastery vs. performance goals.

Finding: SA-CUA is a dual-LLM framework that creates a security boundary for computer-use AI agents by separating trusted and untrusted screen content, preventing indirect prompt injection attacks while reducing user confirmation burden.

Why it matters: Tangentially relevant to AI impact on work and HCI (specifically the "confirmation fatigue" problem in human-AI interaction), but the paper is primarily a security/systems engineering contribution.

Jargon: *Indirect prompt injection (IPI)*: an attack where malicious instructions are embedded in content an AI agent reads, hijacking its actions. *Accessibility tree*: a structured representation of UI elements used by assistive technologies. *Confirmation fatigue*: users becoming desensitized to security prompts and approving them without careful review.

Finding: CoSIR is a post-training method for LLM agents that uses Bayesian information-theoretic rewards to teach multi-agent collaboration behaviors (sharing evidence, requesting information, incorporating teammates' input) that generalize across tasks.

Why it matters: Tangentially relevant to AI impact on collaboration and collective intelligence, but this is primarily a machine learning systems paper with minimal connection to human behavior, HCI, or organizational team dynamics.

Jargon: *Team surprisal*: sum of negative log probabilities each agent assigns to the correct answer given their evidence — a measure of collective uncertainty. *Self-play*: agents train by interacting with copies of themselves.