GenAI-Creativity Meta-Analysis

Differential Effects of Generative AI on Creativity Subfactors: A Three-Level Meta-Analysis

생성형 AI가 창의성 하위요인에 미치는 차별적 효과: 유창성, 융통성, 독창성, 정교성에 대한 3수준 메타분석

---

Research Question

Main RQ: Which aspects of creativity does GenAI enhance or suppress?

AI가 창의성의 어떤 측면을 향상/억제하는가?

Sub-Questions

1. RQ1: What is the effect of GenAI on each subfactor (Fluency, Flexibility, Originality, Elaboration)?
2. RQ2: Are effect sizes significantly different across subfactors?
3. RQ3: How do effects vary by AI type, measurement tool, and task characteristics?

---

Theoretical Framework

Guilford's Divergent Thinking Theory (1967)

Subfactor	Definition	Expected Effect
Fluency (유창성)	Number of relevant responses	Positive (g = 0.35-0.50)
Flexibility (융통성)	Number of different categories	Null/Positive (g = 0.10-0.25)
Originality (독창성)	Statistical rarity of responses	Null/Negative (g = -0.10-0.10)
Elaboration (정교성)	Detail and development	Positive (g = 0.30-0.45)

---

Hypotheses

H1: GenAI → Fluency (+)      [Parallel retrieval mechanism]
H2: GenAI → Flexibility (0/+) [Cross-domain vs. category bias]
H3: GenAI → Originality (0/-) [Mode collapse mechanism]
H4: GenAI → Elaboration (+)   [Autoregressive generation]

H5a: g(Fluency) > g(Originality)
H5b: g(Elaboration) > g(Originality)

---

Methodology

• Design: Three-Level Random-Effects Meta-Analysis
• Effect Size: Hedges' g (small-sample corrected)
• Heterogeneity: I² + Q-test + Prediction Interval
• Publication Bias: Funnel Plot + Egger's Test + Trim-and-Fill
• Software: R (metaSEM, metafor packages)

---

Project Structure

GenAI-Creativity/
├── .research/                    # Research state (Diverga v6.0)
│   ├── project-state.yaml       # Project context & checkpoints
│   └── decision-log.yaml        # Human decision audit trail
├── docs/
│   ├── theoretical-framework/   # Theory documentation
│   ├── protocol/                # PROSPERO protocol
│   └── search-strategy/         # Search documentation
├── data/
│   └── search-results/          # Literature search results
├── scripts/                     # Automation scripts
├── analysis/                    # [Future] R analysis scripts
├── output/                      # [Future] Results & figures
└── README.md

---

Stage Progress

Stage	Name	Status
1	Protocol & Registration	✅ Complete
2	Literature Search	✅ Complete
3	Screening	✅ Complete
4	Data Extraction	🔄 In Progress
5	Quality Assessment	⏳ Pending
6	Meta-Analysis	⏳ Pending
7	Writing	⏳ Pending
8	Publication	⏳ Pending

---

Search Results (Stage 2)

Metric	Value
Databases	OpenAlex, arXiv, PubMed, ERIC
Raw Results	1,338 papers
After Deduplication	1,303 papers
Creativity-Relevant	933 papers
With Abstracts	89.6%
Open Access Rate	87.5%

Papers by Database

Database	Papers
OpenAlex	833
arXiv	370
PubMed	86
ERIC	14

Papers by Year

Year	Papers
2022	38
2023	437
2024	412
2025	379
2026	37

---

Screening Results (Stage 3)

Metric	Value
Screening Tool	Groq API (llama-3.3-70b-versatile)
Total Screened	933 papers
INCLUDE	73 papers (7.8%)
MAYBE	212 papers (22.7%)
EXCLUDE	648 papers (69.5%)
Potential Inclusion Rate	30.5%

Next Step

Full-text review and data extraction from coding Excel file.

---

Data Extraction (Stage 4)

Metric	Value
Total Papers for Review	319
From Screening (INCLUDE)	73
From Screening (MAYBE)	212
Web Search Additions	24
Elicit Suggested	10
Effect Size Rows	1,276 (4 subfactors × studies)

Output Files

• data/coding/GenAI_Creativity_Coding_2026-01-26.xlsx
• data/coding/Elicit_Search_Prompts_2026-01-26.txt

Excel Sheets

1. 0_Summary: Overview statistics
2. 1_Codebook: Variable definitions and coding rules
3. 2_Study_Characteristics: All 319 papers with metadata
4. 3_Effect_Sizes: Template for Fluency, Flexibility, Originality, Elaboration
5. 4_Elicit_Prompts: Search prompts for additional paper discovery

Web Search Sources (Key Papers Added)

• Science Advances - GenAI enhances individual creativity
• Nature Scientific Reports - AI vs humans on divergent thinking
• Journal of Creative Behavior - Human-AI Co-Creativity
• arXiv - GenAI and Creativity Meta-Analysis

Sample Included Papers (Creativity Measures Identified)

• Studies measuring fluency, flexibility, elaboration
• Studies using ChatGPT, GPT-4, generative AI tools
• Experimental and quasi-experimental designs
• Various creativity measurement tools (TTCT, AUT, CAT, etc.)

---

Human-Centered Checkpoints (Diverga v6.0)

This project uses Diverga v6.0 Human-Centered Edition for research coordination.

Approved Checkpoints

Checkpoint	Decision	T-Score	Date
CP_RESEARCH_DIRECTION	하위요인별 차별적 효과	0.35	2026-01-25
CP_THEORY_SELECTION	Guilford (1967) 4요인	0.55	2026-01-25
CP_HYPOTHESIS_TYPE	Directional	-	2026-01-25
CP_METHODOLOGY_APPROVAL	Three-Level Meta-Analysis	0.45	2026-01-26
CP_SEARCH_STRATEGY	Broad Coverage + Comprehensive	0.50	2026-01-26
CP_SCREENING_CRITERIA	Groq + Standard PRISMA	0.50	2026-01-26

T-Score Legend

• T >= 0.7: Common (modal, safe)
• T = 0.4-0.7: Moderate (balanced)
• T = 0.2-0.4: Innovative (differentiated)
• T < 0.2: Experimental (high risk/reward)

---

Key Documents

1. Theoretical Framework
2. PROSPERO Protocol
3. Search Strategy
4. Search Results Summary
5. Decision Log

---

Tools & Frameworks

• Research Coordinator: Diverga v6.0.1
• Analysis: R 4.3+ (metaSEM, metafor, clubSandwich)
• Reference Management: Zotero
• Screening: Rayyan

---

License

[To be determined]

---

Citation

[To be completed upon publication]

---

Generated with Research Coordinator v6.0 Human-Centered Edition