Joshua Holstein, Philipp Spitzer, Sebastian Schäfer et al. (5 authors) • Semantic Scholar
Finding: Feedback interventions targeting retrospective reflection ("reflection-on-action") on AI overrides influence reliance behavior in human-AI teams, but unidirectional feedback can suppress concurrent reflection ("reflection-in-action"), while balanced feedback preserves both without harming overall team performance.
Why it matters: Directly relevant to designing effective human-AI collaboration systems and understanding how feedback design shapes appropriate AI reliance—core concerns for HCI, behavioral science, and AI's impact on work.
Method: Behavioral experiment (N=244) testing feedback intervention conditions on AI override decisions.
Jargon: *Reflection-in-action* = real-time deliberation during a decision; *Reflection-on-action* = retrospective examination of past decisions; *Reliance behavior* = degree to which humans follow or override AI recommendations; *Appropriate reliance* = calibrated agreement with AI that tracks actual AI accuracy.
Alice Rollwagen, E. Schiller, A. Hummel et al. (6 authors) • Semantic Scholar
Finding: Healthcare professionals frequently misinterpreted AI uncertainty communication in nurse scheduling tools, which reduced decision confidence, while AI explanations were perceived positively but didn't significantly affect advice acceptance.
Why it matters: Directly relevant to HCI/UX design and human-AI collaboration topics, particularly how transparency mechanisms (explainability, uncertainty visualization) shape trust and decision-making in AI-assisted workflows.
Method: Mixed-methods user study combining thematic analysis, questionnaire data, and eye tracking with expert users evaluating an iteratively co-designed prototype.
Jargon: *Uncertainty communication* = displaying confidence intervals or probability ranges around AI predictions; *explainable AI (XAI)* = systems that provide human-readable rationales for AI outputs; *decision support system* = software that augments human decision-making rather than replacing it.
Qi Chai, Zhang Zheng, Lu Yang et al. (5 authors) • Semantic Scholar
Finding: PREFER is a framework using LLMs as virtual playtesters to evaluate gameplay experience across psychometric dimensions and guide persona-specific game refinement, achieving measurably better preference ratings than baselines.
Why it matters: Relevant to AI's role in creative/design workflows and games, specifically how LLMs can simulate user feedback to drive iterative design—a potential teaching case for AI-assisted design and human-AI collaboration.
Method: Five personas simulate distinct player profiles; LLMs evaluate games on psychometric experience dimensions adapted from established questionnaires, then iteratively revise rules and playtest candidate games.
Jargon: *Persona*: simulated player profile with distinct preferences used to personalize recommendations; *virtual playtester*: LLM used in place of human playtesters to generate gameplay feedback.
⚠️ Self-reported data only | No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
Ilaria Imperatrice, T. Kameneva, A. Greenberg et al. (4 authors) • Semantic Scholar
Finding: An AI decision model (Jev) designed for fast, structured decisions avoids many classic human cognitive biases at the answer-selection level, but shows significant vulnerabilities to irrelevant numerical anchoring, response ordering effects, and inconsistent probability reasoning across reformulations of the same problem.
Why it matters: Directly relevant to AI impact on work and decision-making, and has implications for behavioral science (cognitive biases, anchoring) when evaluating AI systems deployed in automated decision pipelines.
Method: Large-scale evaluation (41,500+ responses) across cognitive reflection tasks, five cognitive bias paradigms, and demographic pricing experiments—notably testing the conjunction fallacy across 30 task formulations.
Jargon: *System One decision-maker*: AI modeled on fast, intuitive (rather than deliberative) human reasoning; *conjunction fallacy*: the error of rating a specific scenario as more probable than a broader one it's contained within (the classic "Linda problem").
⚠️ No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
C. B. Omeh, Ndudi O. Ezeamuzie, C. Olelewe • Semantic Scholar
Finding: Integrating Google Gemini into problem-based learning improved programming achievement and AI trust among Nigerian undergraduates, with students valuing it for debugging, clarification, and immediate feedback.
Why it matters: Touches on AI impact on learning and HCI, but narrowly focused on a programming education context in a specific regional setting with limited generalizability.
Method: Quasi-experimental pre-test/post-test with intact classes (n=62), ANCOVA with HC3 robust estimation and bias-corrected bootstrapping, plus thematic interview coding.
⚠️ Single study (not replicated) | Convenience sample (e.g., MTurk, students)
Lin Cheng, Jazihan Mahat, M. Puad et al. (4 authors) • Semantic Scholar
Finding: Immersive VR instruction improved primary school students' learning outcomes (achievement, engagement, classroom interaction) for Tang poetry compared to traditional teaching, with the largest effect on engagement.
Why it matters: Marginally relevant to learning sciences and educational technology, but focused narrowly on K-6 Chinese literature instruction rather than HCI, creativity, teams, or behavioral science.
Method: Quasi-experimental ANCOVA design with 166 sixth-graders, using pre-test scores as covariate to control for baseline differences.