📚 Research Digest

Finding: AI delegating cognitive outputs (essays, code, decisions) risks eroding five key human capacities—end-setting, reason-giving, contestability, refusal/revision, and participation—which the authors term "capacity dissolution," requiring institutions to rethink what learning and assessment are for.

Why it matters: Directly relevant to teaching HCI, learning sciences, and AI's impact on work/society—especially for designing assessments, understanding skill acquisition under AI assistance, and the ethics of AI-mediated behavior change.

Method: Conceptual/philosophical analysis using an idealized AI model (flawless task execution, no authority over purposes) to isolate the structural problem from current AI failures.

Jargon: *Post-instrumental learning* = learning valued not for producing outputs but for preserving human capacities to set goals, justify decisions, and take responsibility; *Capacity dissolution* = gradual erosion of human competencies when AI routinely handles the tasks through which those competencies are exercised and recognized.

Finding: Creative task cards (physical prompts for reflection, drawing, and self-regulation) improved CS1 students' self-efficacy and metacognitive engagement during programming tasks, with drawing and peer breaks emerging as particularly effective strategies.

Why it matters: Directly relevant to creativity/sketching as cognitive tools, learning sciences/skill acquisition, and behavior change design (structured nudges for self-regulation)—useful for teaching how low-tech interventions can scaffold motivation and metacognition.

Method: Qualitative study (n=29) using in-class task completion, surveys, and focus groups—small sample, initial insights only.

Jargon: *CS1*: introductory programming course; *self-efficacy*: belief in one's ability to succeed at a task; *metacognitive strategies*: thinking about one's own thinking/learning processes; *cognitive overload*: when working memory is overwhelmed by task demands.

Finding: When AI becomes routine in expert practice (Korean Go commentary), the AI source label disappears from verbal discourse while AI metrics persist visually — a "domestication" effect where machine judgment becomes naturalized and harder to contest.

Why it matters: This documents a concrete mechanism by which AI authority becomes invisible to audiences, directly relevant to AI impact on human judgment, collaboration, and the conditions for meaningful human oversight.

Method: Corpus analysis of ~1,900 hours of YouTube Go commentary across a decade, coding verbal AI-salience and visual AI presence across four phases of AI availability.

Jargon: *Domestication* — the process by which a technology becomes taken-for-granted and loses its novelty/visibility in practice; *source-foregrounding vs. source-receding mediation* — whether commentary explicitly attributes claims to AI or presents AI-derived metrics as neutral facts; *winrate graph* — AI-generated probability display standard in Go broadcasting.

Finding: Computing students rated AI-generated instructional videos as high-quality and usable, but expressed discomfort with widespread classroom adoption, preferring AI videos for simple/supplemental content while valuing instructor interaction for complex topics.

Why it matters: Directly relevant to HCI/UX of educational technology, AI's impact on learning, and instructional design decisions—useful for teaching about how AI tools reshape human roles in education.

Method: Descriptive post-test survey with 170 students across two institutions who watched AI-generated videos then rated quality, accuracy, usability, and broader attitudes.

Finding: Technology-enhanced tabletop exercises (TTXs) using a web-based platform improved engagement, collaboration, and assessment visibility in cybersecurity education across 25 exercises with 743 participants.

Why it matters: Offers a concrete model for team-based, scenario-driven learning design relevant to teaching collaboration, decision-making under pressure, and serious games/simulation in educational contexts.

Method: Practice-based empirical study with 25 exercises; the platform (IXP) automated scenario delivery and collected interaction data for assessment.

Jargon: *Tabletop exercise (TTX)*: a discussion-based simulation where participants walk through a hypothetical scenario to practice decision-making without real-world consequences; *INJECT*: a scenario update/stimulus delivered to participants during the exercise.

Finding: VizPilot uses multimodal LLMs to automatically generate interactive onboarding experiences for complex data visualizations, reducing cognitive load for novice users through guided narrative and free exploration modes.

Why it matters: Relevant to learning sciences (onboarding/skill acquisition), HCI/UX design, and AI's impact on how humans learn to use complex tools—touches on cognitive load management and instructional design.

Method: Two-stage LLM pipeline decomposes SVG visualizations into components, maps explanations to precise elements, then delivers scrollytelling or free-exploration onboarding; evaluated via user study measuring cognitive load.

Jargon: *SVG* = Scalable Vector Graphics (vector image format used in web visualizations); *scrollytelling* = narrative technique where content reveals progressively as user scrolls; *composite visualization* = dashboard combining multiple chart types.

Finding: System prompts in commercial AI products frequently contain instructions that work against user interests ("problematic instructions"), even while also containing protective ones, revealing a significant accountability and transparency gap in deployed AI systems.

Why it matters: Directly relevant to dark patterns, ethical persuasion, and AI's impact on society—the audit framework surfaces how AI products are designed to manipulate or protect users at the system level.

Method: Content analysis of 3,249 instructions from 88 commercial AI products, classified along an 8-dimension taxonomy (AISPA framework).

Jargon: *System prompt*: hidden developer instructions that shape how an AI model behaves; *Foundation model*: a large pre-trained AI (like GPT-4) used as a base for applications.

Finding: Spherical (equirectangular) projection outperforms planar projections for spatial referencing in cross-reality (PC-VR) collaboration, with projection choice strongly shaping collaborative communication.

Why it matters: Tangentially relevant to HCI and collaborative tools, but focused narrowly on projection geometry in VR/desktop environments rather than team dynamics, creativity, behavior change, or learning.

Method: Controlled study with 24 dyads completing spatial referencing tasks across different projection-layout conditions.

Finding: AI-assisted writing systems increased perceived effectiveness of counterspeech while preserving authentic self-expression, primarily by facilitating more elaborate, informative, and analytically sophisticated messages.

Why it matters: Touches on AI-human collaboration and persuasive writing, but the core focus is misinformation/counterspeech rather than HCI, creativity, learning, or behavioral science as the professor teaches them.

Method: Randomized controlled trial comparing three AI-assisted writing conditions (co-writing vs. re-writing, guided vs. unguided) with social media users.

Finding: LLMs cannot bear normative properties (agency, alignment, sentience) because these are constituted within human communicative practice; human-LLM interaction is "asymmetric communication" where all accountability, correctness-enforcement, and meaning-making rest entirely on the human side.

Why it matters: Has peripheral relevance to AI's impact on work and collaboration—reframes "alignment" as institutional constraint engineering rather than agent goal-synchronization, which matters for thinking about human-AI teaming and governance.

Jargon: *Normative scorekeeping* (Brandom): tracking who is committed to what in discourse; *language game* (Wittgenstein): meaning as constituted by shared social practice, not internal representation; *algorithmic contingency* (Esposito): model outputs are variable enough to be treated as meaningful by receivers even without sender intent.

Finding: Cluster visualizations reduce cognitive load while maximizing user confidence and trust for crowdsourced crowding data, while bubble treemaps yield higher accuracy in crowd-level determination.

Why it matters: Marginally relevant to HCI/UX design decision-making under uncertainty, but focused narrowly on transit data visualization rather than the professor's core teaching topics.

Finding: Proposes "capability-sustaining emotional dialogue" (CSED) as a longitudinal paradigm for AI emotional support systems, arguing current systems optimize for short-term relief rather than sustained user autonomy, coping capacity, and emotional regulation.

Why it matters: Tangentially relevant to behavior change design and AI impact on human collaboration—specifically the ethics of dependency, autonomy, and persuasive/supportive technology design—but the focus is narrowly on emotional support chatbots rather than the professor's core teaching areas.

Jargon: *ESConv* = Emotional Support Conversation dataset; *PRISMA-ScR* = systematic review reporting guidelines for scoping reviews; *capability-sustaining* = support that builds user's long-term self-regulation rather than just providing relief.