A. C. Jacomuzzi, Brigitta Pia Alioto, Sara Magni • Semantic Scholar
Finding: This scoping review of 10 review-level publications maps how cognitive biases are conceptualized and addressed in human-AI interaction, proposing "coupled human-AI cognition" as a theoretical lens grounded in distributed cognition and joint cognitive systems frameworks.
Why it matters: Directly relevant to teaching behavioral science (cognitive biases in decision-making) and AI's impact on human collaboration, particularly how AI systems can amplify or transform cognitive biases in high-stakes domains.
Method: Structured conceptual mapping review across Scopus, Web of Science, and PubMed (n=10 eligible review publications, 2020–2025), applying a three-level framework: data-level, algorithmic-level, and human-interaction-level.
Jargon: *Coupled human-AI cognition* — the idea that human and AI reasoning are interdependent rather than separate; *joint cognitive systems* — framework treating human-machine pairs as a single cognitive unit; *distributed cognition* — cognition is spread across people, artifacts, and environments rather than residing solely in individual minds.
Yan-Ming Hou, Ying-Ying Zhang, Bin Wang et al. (5 authors) • Semantic Scholar
Finding: Human evaluators show domain-specific bias in AI creativity attribution—favoring human-labeled outputs in artistic domains (painting, poetry) but rating identical outputs as *more* creative when labeled AI in divergent thinking tasks (Alternative Uses Test).
Why it matters: Directly relevant to teaching about AI's impact on creative work, human-AI collaboration, and cognitive biases in evaluation—useful for HCI, creativity, and behavioral science courses discussing how source attribution shapes perception.
Method: Three experiments using attribution manipulation (same outputs labeled "AI" vs. "human" vs. co-created) across painting, poetry, and divergent thinking tasks.
Jargon: *Alternative Uses Test* — a standard divergent thinking measure where participants generate novel uses for everyday objects; *attribution bias* — crediting outcomes differently based on perceived source rather than actual content.
⭐ Includes replication ⚠️ No control group mentioned | Limited statistical detail in abstract
Finding: Adults develop personalized "good digital practices" as relational competencies rather than universal norms, navigating tensions between digital connection and overstimulation/loss of control through reflection and shared discourse.
Why it matters: Directly relevant to behavior change design, persuasive technology ethics, and attention/habit formation in digital environments — especially the tension between engagement mechanisms and user autonomy/well-being.
Method: Longitudinal qualitative study with 100 adults using a "digital community" method (journals, forums, social space across four waves) analyzed via discourse analysis — the digital community methodology itself is presented as a novel contribution for studying automated everyday behaviors.
Jargon: *Digital ecologies* = interconnected systems of platforms, algorithms, and social practices shaping everyday life; *discursive positions* = distinct stances participants adopt when making sense of and narrating their digital lives.
⭐ Longitudinal design ⚠️ No control group mentioned | Single study (not replicated)
Xiao-Long Li, Xiao-Han Xu, Jin-Yang Li et al. (8 authors) • Semantic Scholar
Finding: GenUI-Harness is a multi-agent system that generates ephemeral UI interfaces to replace text-based human-agent interaction, reducing dialogue rounds from 3.4 to 1.2 and improving task completion in database-backed workflows.
Why it matters: Directly relevant to HCI and UX design — the core argument that structured generative UIs reduce cognitive overload over natural language interfaces connects to interaction design principles and AI's impact on human-computer collaboration.
Method: Reinforcement learning with a novel "Reward Auditor" meta-reward mechanism and a new benchmark (UI-TAU Bench) across 300–1,000 real-world tasks.
Jargon: *Generative UI* — interfaces dynamically created by AI at runtime rather than pre-designed; *Pass@3* — whether a correct solution appears within 3 attempts; *reward hacking* — when a model exploits flaws in the reward signal rather than genuinely improving; *LLM-as-a-Judge* — using a language model to evaluate outputs as a proxy for human judgment.
Finding: Active pop-up annotations boost immediate vocabulary recognition but hurt deep reading comprehension, while inline glosses offer a middle-ground trade-off; this frames annotation design as an HCI decision with measurable learning consequences.
Why it matters: Directly relevant to HCI design decisions and learning sciences, specifically how interaction modality (active vs. passive) affects cognitive load and skill acquisition in language learning contexts.
Method: Between-subjects experiment (n=90) comparing three conditions on mobile devices, measuring both vocabulary recognition and reading comprehension scores.
Jargon: *Inline glosses* = definitions displayed directly within the text without user action; *click-to-reveal pop-ups* = definitions hidden until the user taps a word (active interaction).
Lussy Midani Rizki, Y. S. Kusumah, Al Jupri et al. (7 authors) • Semantic Scholar
Finding: Calcyon is an AI chatbot framework for university calculus that integrates computational thinking scaffolding and self-regulated learning principles, validated via expert review and a small pilot study.
Why it matters: Touches on AI in learning and self-regulated learning design, but narrowly focused on math education tool development with no effectiveness data—limited generalizability to HCI, creativity, or behavioral science teaching contexts.
Method: Design and Development Research (DDR) using ADDIE model; expert validity via Aiken's V coefficient.
Jargon: *Aiken's V* — a statistical measure of content validity based on expert ratings; *ADDIE* — Analyze, Design, Develop, Implement, Evaluate instructional design model; *SRL* — self-regulated learning (metacognitive goal-setting and monitoring of one's own learning).
⚠️ Very small sample (n=11) | No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
Finding: LLM agents learn better from retaining complete experience records than from summarizing rules, and still underperform humans who explore more varied strategies with less repetition.
Why it matters: Marginally relevant to AI's capabilities as collaborative/learning agents, but the focus is narrowly on benchmark engineering for game-playing LLMs rather than HCI, creativity, teams, or behavior change.
Method: Novel text-based games with counterintuitive rules designed to prevent models from relying on pretrained knowledge, with repeated attempts and varied instances.
Max T. C. Wong, Kai-Tik Lee, Yuan-Rou Chen et al. (14 authors) • Semantic Scholar
Finding: Curriculum-aligned AR app with 3D models produces significantly larger conceptual gains (d=0.32–0.48) and higher engagement than textbook images for secondary physics students learning circular motion and magnetic force.
Why it matters: Marginal relevance — touches on learning sciences and educational technology, but focused narrowly on physics content delivery rather than HCI design, creativity, collaboration, or behavioral science.
Method: Quasi-experimental crossover design with N=171 students across 9 classes, comparing AR vs. textbook with pre/post assessments and structured behavioral observation.
⚠️ Self-reported data only | No control group mentioned | Single study (not replicated)
Finding: Proposes a three-tier architecture for deploying generative AI in mobile learning, using cognitive load-aware routing and dynamic human-machine role switching to balance computational constraints with instructional effectiveness.
Why it matters: Touches on AI in learning and adaptive human-machine collaboration, but the focus is overwhelmingly on systems engineering (routing protocols, pruning, cache optimization) rather than HCI, pedagogy, or behavioral outcomes.
Jargon: *Cognitive load routing* = directing computation based on mental demand of tasks; *local clustering pruning* = reducing model size for on-device efficiency.
Shan-Zhen Gao, Wei-Zheng Gao, Julian D. Allagan et al. (5 authors) • Semantic Scholar
Finding: Proposes a human-centered pedagogical framework for teaching blockchain, FinTech, and generative AI to HBCU undergraduates using active learning, structured collaboration, role rotation, and peer evaluation.
Why it matters: Has marginal relevance to learning sciences and AI's impact on education, but the core focus is curriculum/pedagogy design for a specific institutional context rather than HCI, creativity, team dynamics, or behavioral science.