Finding: Online communities collectively negotiate trust and mistrust in generative AI through distinct discourse patterns—trust clusters around companionship/emotional support while mistrust focuses on technical failures, jailbreaks, and inaccuracy—with emotional trust driving higher engagement than mistrust.
Why it matters: Directly relevant to AI's impact on work/learning/creativity and behavioral science around how people form attitudes toward AI tools in social contexts.
Method: Structural topic modeling on 301,498 Reddit posts combined with a theory-guided trust/mistrust dictionary and regression analysis—a strong computational social science approach for understanding collective sensemaking at scale.
Jargon: *Structural topic modeling* = a method that identifies latent topics in text corpora while allowing topic prevalence to vary with metadata (time, community); *Sycophancy* = AI tendency to agree with users rather than provide accurate responses.
Finding: Reddit communities negotiate trust and mistrust in generative AI collectively, with trust centered on companionship/emotional support and mistrust on jailbreaks, degradation, bugs, and inaccurate answers; emotional trust content drives higher engagement while mistrust topics suppress it.
Why it matters: Directly relevant to AI's impact on work, learning, and creativity, and provides empirical grounding for how users collectively make sense of AI tools—useful for teaching AI literacy, HCI, and behavioral science around trust.
Method: Structural topic modeling on 301,498 Reddit posts combined with a theory-guided trust/mistrust dictionary and regression analyses on engagement metrics.
Jargon: *Structural topic modeling* — a method that discovers latent topics in text while incorporating document-level covariates (e.g., subreddit, time); *sycophancy* — AI tendency to agree with or flatter users rather than provide accurate responses.
Shayla Sharmin, Mohammad Fahim Abrar, Mohammad Al-Ratrout et al. (4 authors) • ArXiv
Finding: Two LLM-powered scaffolding systems (indirect via guiding questions vs. direct via worked examples) both supported student problem-posing refinement, with direct scaffolding producing faster improvements but indirect scaffolding promoting deeper reflection; authors recommend sequencing indirect before direct.
Why it matters: Directly relevant to learning sciences and AI-in-education — demonstrates how AI scaffolding design choices shape metacognitive engagement and skill acquisition, with practical implications for HCI/UX design of educational AI tools.
Method: Within-subjects counterbalanced pilot study (n=20 graduate students) using Bloom's Taxonomy-based automated assessment, combining problem-quality ratings, UX surveys, and post-session interviews.
Jargon: *Problem posing* — learners generate and refine their own problems rather than solving given ones; *scaffolding modality* — the form of support provided (questions vs. examples); *Bloom's Taxonomy* — hierarchical framework of cognitive learning objectives used here as assessment criteria.
⭐ Longitudinal design ⚠️ Self-reported data only | No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
Finding: Experienced developers want significantly more configurability in AI coding assistants than current tools provide, particularly for task-related controls (confidence thresholds, suggestion quality visibility), but find persona-related configurations less useful.
Why it matters: Directly relevant to HCI/UX design principles for AI tools and raises important questions about human agency and control in AI-augmented work environments.
Method: Mixed-methods: survey of 56 professional developers plus 7 design sessions where participants built their ideal "control board" for AI assistants.
Jargon: *Minimum confidence threshold* = a setting that only shows AI suggestions when the AI's internal certainty exceeds a specified level; *configuration surface* = the interface layer through which users access and adjust system settings.
⚠️ Self-reported data only | No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
Sophia Liu, Shm Garanganao Almeda, Max Kreminski et al. (4 authors) • ArXiv
Finding: UISTful is an interactive system that transforms individual reading activity into collective visualizations of a scholarly community, treating reading traces as creative, authorial acts inspired by the concept of the "information flâneur."
Why it matters: This is directly relevant to HCI (it's a UIST paper on a novel reading/exploration interface), touches on collective intelligence and community sense-making, and raises interesting questions about how shared attention data can reveal community knowledge structures.
Method: System built around a semantic globe visualization of papers/authors, with private reading traces that users can curate and publish for community replay.
Jargon: *Information flâneur* — borrowed from Baudelaire's concept of an urban wanderer; here applied to a reader who drifts through information spaces without a fixed agenda. *Reading trace* — a recorded log of a user's navigation path through a document collection.
Stefano Balietti, Pietro Saggese, Markus Strohmaier • ArXiv
Finding: In DAO governance, voting outcomes are systematically skewed by author endorsement (+58.8%), approval-oriented framing (+27.1%), and first-position placement (+7.7%), suggesting these interface and social design choices function as de facto institutional biases.
Why it matters: Directly relevant to behavioral science (order effects, anchoring, social influence) and to HCI/UX design—the paper reframes interface choices in voting systems as non-neutral design decisions with measurable behavioral consequences.
Method: Proposal-choice level regression linking voting-power share to three observable features across multiple DAO governance datasets, with robustness checks subtracting author's own voting power.
Jargon: *Token-weighted voting* = voting power proportional to cryptocurrency tokens held; *DAO* = blockchain-based organization governed by smart contracts and member votes; *position/order effects* = cognitive bias favoring items listed first.
Jingbo Chen, He Wang, Wei Yuan et al. (5 authors) • ArXiv
Finding: AutoWorldBuilder is a multi-agent LLM system for fictional worldbuilding that achieves 95% success rate across 20 tasks by combining context compression, conflict detection, and specialized reviewer agents to generate self-consistent fictional worlds.
Why it matters: Directly relevant to game design and creative AI tools, and the multi-agent collaboration architecture has implications for how AI systems support creative work and team-like division of labor in creative tasks.
Method: Two experiments across 20 worldbuilding tasks using two LLM backends (GPT-OSS 120B, DeepSeek v3.2), measuring success rate, concept count, and consistency.
Jargon: *DAG-based scheduler* = Directed Acyclic Graph used to order tasks by dependencies; *context explosion* = the problem of LLM context windows filling up as conversation history grows; *temperature configuration* = parameter controlling LLM output randomness/creativity.
⚠️ No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
Gwenn Beets, Anniek Jansen, Saar Hommes et al. (9 authors) • ArXiv
Finding: Perceived benefits drive intention to use AI health chatbots and willingness to self-disclose, while perceived risks suppress both; topic type (physical vs. psychological) matters less than individual characteristics and topic sensitivity.
Why it matters: Directly relevant to the "Friend or Therapist?" chatbot review the professor found relevant, and informs HCI design of AI conversational agents—particularly around trust, disclosure, and sensitive topic handling.
Method: Mixed-design online experiment (between/within subjects, N=1,388, Dutch representative sample) manipulating topic type and sensitivity.
Jargon: *Self-disclosure* = voluntarily sharing personal information with another party (here, an AI chatbot); *topic sensitivity* = perceived social risk or stigma associated with revealing a particular health topic.
⭐ Longitudinal design ⚠️ Single study (not replicated) | Limited statistical detail in abstract
Finding: Small hyperbolic language models (146M–3B parameters) can detect sycophancy, dependence-fostering, and false memories in AI companions, while implementing "designed forgetting" and creativity—proposed as a route to trustworthy companion AI.
Why it matters: Tangentially relevant to AI's impact on human relationships and the chatbot-as-companion design problem, which connects to the user's interest in AI's societal effects and the chatbot/therapist literature they found relevant.
Method: Pairwise comparison study (n=311 decisions) for creativity evaluation; AUROC-based leave-one-generator-out evaluation for sycophancy/compliance detection.
Jargon: *Hyperbolic substrate* = embedding space with hyperbolic geometry, useful for hierarchical representations; *confabulated memories* = AI-generated false memories presented as real; *skeleton-wallpaper partition* = a memory architecture distinguishing stable core facts from contextual details; *Fleiss kappa* = inter-rater reliability statistic.
Finding: Machine learning models (tree-based ensembles) can predict monster difficulty levels in tabletop RPGs from numerical attributes with high accuracy, potentially serving as a computer-aided design tool for game designers.
Why it matters: Marginally relevant to game design and AI-assisted creative work, but the focus is narrowly on ML methodology rather than HCI, design process, or human factors.
Method: Ordinal regression with chronological/expanding-window evaluation protocols to simulate real design workflows.
Elizabeth L Fisher, Jakob Hohwy, Claire O'Callaghan • PsyArXiv
Finding: Proposes reconceptualizing optimism bias as a naturalistic, evolutionarily conserved behavior rather than a deliberative judgment, linking human and animal research through an ecological framework.
Why it matters: Optimism bias is tangentially relevant to goal-setting, motivation, and behavioral science, but this paper is primarily an evolutionary/neuroscience framework paper with no application to HCI, teams, learning, or AI.