Muhammad Muaavia Khalid, Binish Nawaz, Faiza Abdul Latif et al. (4 authors) • Semantic Scholar
Finding: Perceived "understanding" in AI conversations is a psychologically constructed experience driven by linguistic alignment, contextual coherence, personalization, response timing, and anthropomorphic design cues—not actual machine comprehension.
Why it matters: Directly relevant to behavior change design and persuasive technology ethics, particularly around AI-induced emotional over-attribution, dark patterns, and user engagement mechanisms that exploit psychological vulnerabilities.
Method: Conceptual mixed-methods using secondary empirical literature synthesis and modeled perceived understanding as a function of multiple interactional variables.
Jargon: *Linguistic alignment* = users and AI mirroring each other's language style; *anthropomorphic cues* = design features that make AI seem human-like; *social presence* = feeling of interacting with a real person.
Finding: Physicians showed partial, strategic reliance on LLM-generated psychiatric assessments, with performance gains inversely related to baseline competence—less confident users relied more on AI, but miscalibrated trust (over-skepticism or overreliance) reduced gains.
Why it matters: The trust calibration and algorithm aversion dynamics are directly relevant to behavioral science and AI-human collaboration research, particularly for teaching how humans decide when to defer to AI systems.
Method: 12 psychiatrists evaluated 50 case vignettes before and after seeing GPT-4o ratings; accuracy measured via weighted kappa against expert consensus gold standard.
Jargon: *Algorithm aversion* = tendency to distrust algorithmic recommendations even when they outperform humans; *switch rate* = proportion of cases where physicians revised their assessment after seeing AI disagreement; *weighted kappa* = inter-rater reliability metric that accounts for degree of disagreement.
Finding: Argues that AI agent design must account for multi-agent ecosystems where misaligned incentives produce poor social outcomes, proposing "Artificial Social Intelligence" as a research agenda covering alignment, fairness, and human-AI coexistence.
Why it matters: The human-AI collaboration and societal impact framing is tangentially relevant to teaching AI's impact on work and society, but stays at a high theoretical level without empirical grounding or direct HCI/behavioral application.
Finding: Consumer neurotechnology systems claiming to detect emotions from facial expressions, voice, physiological signals, or EEG are scientifically unreliable due to fundamental conceptual, empirical, and technical limitations.
Why it matters: Marginally relevant to HCI and behavioral science as a cautionary critique of AI-based emotion recognition tools increasingly used in education, HR, and marketing contexts the professor covers.
Gerard Chung, Lim Tse Min, Joan R. Fang • PsyArXiv
Finding: Chat-based mental health workers for youth primarily use social support, problem-solving, and reframing techniques; these core skills are shared across workers rather than being individual styles, and a small subset predicts better user outcomes.
Why it matters: Tangentially relevant to behavior change design and persuasive technology, but focused narrowly on mental health counseling techniques rather than HCI, UX, or the professor's core teaching areas.
Method: Coded 44,916 worker turns from 1,439 real chat sessions using the Behavior Change Technique Taxonomy v1 (BCTTv1), a standardized 93-item taxonomy for classifying intervention components.
Jargon: *BCTTv1*: Behavior Change Technique Taxonomy version 1, a standardized framework for categorizing discrete active ingredients in behavior change interventions.
Finding: Develops and validates a 27-item School Curiosity Inventory with seven dimensions across motivation and meta-curiosity domains, showing these dimensions predict learning and well-being outcomes better than general curiosity measures.
Why it matters: Marginally relevant to learning sciences and motivation, but focused on psychometric scale development in Chinese K-12 contexts rather than HCI, design, teams, or behavior change.
⭐ Includes replication | ⭐ Large sample size ⚠️ Self-reported data only | Convenience sample (e.g., MTurk, students)
Finding: TEXA OS is a prototype agentic AI platform that combines natural-language understanding, autonomous task planning, and multi-agent execution to automate real-world computer tasks (browser automation, document generation, OS control) rather than just answering questions.
Why it matters: Marginally relevant as an assistive HCI system for accessibility (elderly users, people with disabilities), but this is primarily a systems engineering paper with little behavioral or user-centered analysis.
Method: System evaluation via unit, integration, and system-level tests—not a user study, no behavioral or UX metrics reported.