How Different GenAI Feedback Types Reshape Human-GenAI Interactions in Higher Education: Insights from a Meta-Analysis

This dataset contains the coded variables and effect size data extracted from 52 independent studies investigating the effects of Generative AI (GenAI) feedback on learning outcomes in higher education settings. The data were collected through systematic literature searches across major academic databases (WOS, Scopus, CNKI) and screened according to PRISMA guidelines.Variables Included:S1. Intervention Duration: Categorized as 10 weeks (k=7), representing the length of GenAI feedback intervention in each study.S2. Type of Course: Classified into STEM (programming, mathematics, science; k=22), LAW (English, academic writing, research methodology; k=18), or SCD (teacher training, creative problem-solving, arts design; k=12).S3. Sample Size: Operationalized as Small (1-50 participants; k=15), Medium (51-100; k=24), or Large (>100; k=13).S4. Interaction Mode: Distinguishing between Student-Machine (S-M; direct learner-GenAI interaction; k=31) and Teacher-Student-Machine (S-T-M; triadic collaboration involving instructor mediation; k=21).S5. GenAI Interaction Proficiency: Participant expertise levels coded as Novice (first-time users; k=19), Intermediate (some prior experience; k=23), or Proficient (regular users; k=10).S6. Learning Model: Instructional approaches including Personalized Learning (PL; k=16), Problem-Based Learning (ProbL; k=14), Project-Based Learning (PBL; k=13), and Reflective Learning (RL; k=9).S7. Interactive Rounds: Turn-taking patterns categorized as Single Round (SR; one-shot Q&A without memory; k=11), Multiple Rounds-Instant (MRI; ≥2 turns with ≤30s intervals; k=28), or Multiple Rounds-Delay (MRD; ≥2 turns with >30s intervals; k=13).S8. Interaction Feedback Mode: Communication channels including Text-Audio (unimodal text/voice; k=20), Text-Visual (text plus static/dynamic images; k=19), or Multimodal (AR/VR/embodied combinations; k=13).S9. Interaction Control Methods: Interface designs ranging from Structured Prompts (template-based; k=17), Menu-Driven (limited options; k=8), to Free Dialogue (open-domain natural language; k=27).S10. Types of Dialogue Feedback Purposes: Functional roles comprising Diagnostic/Corrective (DC; error correction; k=21), Coaching/Scaffolding (CS; guided support; k=18), Co-creation/Exploration (CE; collaborative ideation; k=9), or Emotional/Motivational Support (EMS; affective support; k=4).S11. AI Usage Context: Visibility settings coded as Private Use (individual only; k=25), Group Shared (within learning teams; k=18), or Public Shared (classroom-visible; k=9).S12. Type of GenAI Agent: Embodiment status distinguishing Embodied Agents (virtual humans/robots/holograms with nonverbal cues; k=14) from Abstract Agents (text/voice-only without anthropomorphic presence; k=38).Outcome Measures: Effect sizes (Hedges' g) extracted for four primary constructs: Behavioral Intention (BI), Cognitive Development (CD), Emotional and Attitudinal Impact (EAI), and Interpersonal Communication (IPC).Data Format: Excel/CSV format with 63 effect sizes nested within 52 studies, including study identifiers, publication years, sample characteristics (EG/CG means, SDs, N), and quality assessment indicators.