File size: 8,701 Bytes
5fab7c8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 | ---
license: cc-by-4.0
task_categories:
- text-classification
- tabular-classification
language:
- en
tags:
- ai-fitness
- ai-washing
- mobile-apps
- google-play
- health-fitness
- annotation
size_categories:
- n<1k
---
# AI-Fit-Scan: A Labeled Dataset of AI-Powered Fitness Applications on Google Play
## Dataset Description
- **Homepage:** https://huggingface.co/datasets/MaxGuo/ai-fit-scan
- **Repository:** https://github.com/GBX-Max1220/AI-Fit-Scan
- **Paper:** *Coming soon*
- **Point of Contact:** Max Guo (gbx1220max@gmail.com)
### Summary
AI-Fit-Scan is a manually annotated dataset of **161 fitness-related mobile applications** scraped from Google Play, with **61 apps claiming AI capabilities** receiving fine-grained human annotation for AI authenticity. We introduce a three-tier labeling framework (True AI / Quasi-AI / Fake AI) and achieve **inter-annotator agreement of κ = 0.856 (Cohen's Kappa)**, indicating almost perfect agreement.
**Key finding:** Among 61 apps claiming to be "AI-powered fitness" applications, **36.1% (22) are fake AI** — their core functionality does not use AI/ML, and **only 57.4% (35) are confirmed true AI**.
### Why This Dataset Matters
1. **AI-Washing quantification:** First systematic measurement of AI claim inflation in the fitness app market
2. **Reusable annotation framework:** Three-tier (T/Q/F) labeling methodology with proven reliability
3. **AI function taxonomy:** Six-category classification of AI capabilities in fitness apps
4. **Academic utility:** Benchmark for studying AI claim verification, consumer deception, and health app regulation
---
## Dataset Structure
### Files
| File | Description | Rows |
|------|-------------|------|
| `ai_fit_scan_full.csv` | Complete 161-app dataset with L1 metadata labels | 161 |
| `ai_fit_scan_annotated.csv` | 61 AI-claiming apps with L2 human annotation | 61 |
### Columns
**Both files:**
- `name` — App name
- `category` — Google Play category
- `installs` — Download count
- `rating` — User rating (0-5)
- `reviews` — Number of user reviews
- `description` — App description from Google Play
- `l1_label` — L1 automatic classification (AI_FITNESS / FITNESS_NO_AI / EXCLUDE_GENERIC / AI_NOT_FITNESS / UNRELATED)
**Annotated file only:**
- `l2_label` — L2 human annotation (T / Q / F)
- `ai_function` — AI capability category (see taxonomy below)
- `annotator_a` — Annotator A label
- `annotator_b` — Annotator B label
- `l2_evidence` — Evidence for L2 label (source URL or rationale)
### Label Definitions
#### L1: Metadata-Based Classification
| Label | Definition | Count |
|-------|-----------|-------|
| AI_FITNESS | Health/Fitness category + AI keywords in description | 61 |
| FITNESS_NO_AI | Health/Fitness category, no AI keywords | 50 |
| EXCLUDE_GENERIC | Generic AI tools (chatbots, translators, etc.) | 25 |
| AI_NOT_FITNESS | AI keywords but non-fitness category | 13 |
| UNRELATED | Neither fitness nor AI | 12 |
#### L2: Human Annotation of AI Authenticity
| Label | Definition | Count | % |
|-------|-----------|-------|---|
| **T (True AI)** | AI/ML is the core driver of the app's primary functionality. Removing AI would fundamentally change the product. | 35 | 57.4% |
| **Q (Quasi-AI)** | App claims AI but evidence is inconclusive. May use rule-based algorithms or simple heuristics marketed as AI. | 4 | 6.6% |
| **F (Fake AI)** | AI label is marketing only. Core functionality works without AI, or "AI" refers to basic automation/features unrelated to ML. | 22 | 36.1% |
**Annotation criteria:**
- **T:** Description explicitly describes AI/ML-driven personalization, adaptive planning, computer vision, or LLM-based coaching as a core feature
- **Q:** App name or description claims "AI" but lacks specific mechanism description; could be rule-based
- **F:** "Smart"/"AI" used as marketing buzzword; core is workout logging, pre-set routines, or timer functionality; AI only used for peripheral features (e.g., food photo recognition in a calorie counter)
### AI Function Taxonomy
| Category | Description | Count |
|----------|-------------|-------|
| Plan Generation | AI generates/adapts personalized workout plans based on user data | 28 |
| LLM Chat Coach | LLM-powered conversational coaching interface | 8 |
| Nutrition AI | AI-driven diet and nutrition recommendations | 6 |
| Pose/Motion Detection | Computer vision for real-time form checking and rep counting | 5 |
| Wearable Integration | AI adjusts plans based on biometric data (HRV, sleep, etc.) | 4 |
| Voice Coach | Real-time AI voice guidance during workouts | 3 |
---
## Data Collection
### Methodology
1. **Search strategy:** 10 keyword queries on Google Play (`"AI fitness"`, `"AI workout"`, `"AI personal trainer"`, `"AI gym"`, `"AI exercise"`, `"AI coach"`, `"AI training plan"`, `"AI running"`, `"smart fitness"`, `"AI health"`)
2. **Deduplication:** Removed duplicate entries by package name
3. **Collection date:** May 2026
4. **Data extracted:** App name, category, installs, rating, reviews, description, developer, price, version history
### Annotation Process
1. **L1 automatic classification:** Keyword-based filtering using AI-related terms and category matching
2. **L2 human annotation:** Two annotators independently labeled all 61 AI_FITNESS apps
- Annotator A: AI research assistant (automated labeling + L2 web search verification)
- Annotator B: Domain expert (CSCS-certified, HCI researcher)
3. **L2 evidence collection:** For ambiguous cases (initially labeled Q), web search was conducted to verify AI technology claims against official websites, technical documentation, and third-party reviews
4. **Disagreement resolution:** 5 disagreements (all Q vs T/F) resolved by adopting Annotator B's judgment after reviewing app screenshots and user reviews
### Inter-Annotator Agreement
| Metric | Value |
|--------|-------|
| Raw agreement | 56/61 = 91.8% |
| Cohen's Kappa | **0.856** |
| Interpretation | Almost Perfect (Landis & Koch, 1977) |
---
## Key Findings
### 1. AI-Washing is Rampant
36.1% of apps claiming "AI" in the fitness space have no meaningful AI in their core functionality. The most common fake-AI patterns:
- "Smart" = basic algorithm/timer (e.g., Smart Workout Counter = interval timer)
- "AI-Powered" added to existing traditional apps (e.g., MyFitnessPal adding AI food recognition)
- "Personal Trainer" marketing without any ML (e.g., Home Workout - No Equipment)
### 2. Fake AI Apps Have Higher Ratings
| Group | Mean Rating | Median Rating |
|-------|-------------|---------------|
| True AI (T) | 4.28 | 4.50 |
| Fake AI (F) | 4.41 | 4.54 |
**High-rating apps sell outcome promises; low-rating apps sell AI technology.**
### 3. Long Tail of Irrelevance
- 28 of 61 "AI fitness" apps have <50K downloads
- 15 apps have zero or missing ratings — likely inactive or abandoned
- The market has not consolidated; room for genuine AI entrants
### 4. AI Function Distribution
Plan generation dominates (28/35 true-AI apps), while pose detection (5) and voice coaching (3) remain underdeveloped despite high user value potential.
---
## Limitations
1. **Single platform:** Google Play only; Apple App Store data not included
2. **Single time point:** Data collected May 2026; AI claims may change with updates
3. **Description-dependent:** L2 annotation relies primarily on app descriptions and public documentation, not source code inspection
4. **Binary framing:** The T/Q/F taxonomy simplifies a spectrum; some "Quasi-AI" apps may use sophisticated rule engines that approach ML-level personalization
5. **Keyword bias:** Search results favor apps with "AI" in their name/description; genuine AI apps using different marketing language may be missed
---
## Ethical Considerations
- App names and descriptions are publicly available Google Play metadata
- Developer contact and revenue data are **not** included
- We do not accuse any app of fraud; "Fake AI" refers to absence of ML in core functionality, not malicious intent
- Apps labeled F may use AI in peripheral features (e.g., food recognition, recommendation systems)
---
## Citation
```bibtex
@dataset{guo2026aifitscan,
title={AI-Fit-Scan: A Labeled Dataset of AI-Powered Fitness Applications on Google Play},
author={Guo, Baixin (Max) and MaxCoze},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/datasets/MaxGuo/ai-fit-scan}
}
```
## License
Creative Commons Attribution 4.0 International (CC BY 4.0)
## Acknowledgments
- Cal Dietz Triphasic Training system — conceptual influence on training periodization analysis
- Google Play Store — public data source
|