README.md

---
license: cc-by-4.0
pretty_name: "Debt Behind the AI Boom"
language:
  - en
tags:
  - code
  - software-engineering
  - technical-debt
  - ai-generated-code
  - empirical-study
  - static-analysis
  - github
size_categories:
  - 100K<n<1M
configs:
  - config_name: repos
    data_files: ai_repos.csv
---

# Debt Behind the AI Boom — Replication Data

Data for the paper:

> **Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild**
> Yue Liu, Ratnadira Widyasari, Yanjie Zhao, Ivana Clairine Irsan, Junkai Chen, David Lo
> 📄 [arXiv:2603.28592](https://arxiv.org/abs/2603.28592) · 💻 Code: [github.com/yueyueL/tech-debt-ai-coding](https://github.com/yueyueL/tech-debt-ai-coding)

We mined **302.6K AI-authored commits** from **6,299 GitHub repositories** across five
AI coding assistants (GitHub Copilot, Claude, Cursor, Gemini, Devin), ran static
analysis before and after each commit, and tracked whether the introduced issues
still survive in the codebase today.

**Key finding:** 484K issues were introduced; **22.7% still survive at HEAD**, including
issues introduced more than nine months earlier.

> This Hugging Face repository hosts the **data bundles**. All **code** lives on GitHub:
> 👉 **https://github.com/yueyueL/tech-debt-ai-coding**

---

## What's in this dataset

| File | Size | Uncompressed | What it is |
|---|---|---|---|
| `focused_repos.json` | 1.6 MB | — | Manifest of the **6,299 analyzed repositories** (stars, language, AI-commit counts per repo). The authoritative input to the analysis pipeline. |
| `ai_repos.csv` | 0.55 MB | — | Repo metadata (stars, language, primary AI tool) for the same 6,299 repos. *(Browsable in the Dataset Viewer above.)* |
| `commits.zip` | 48 MB | ~265 MB | All 6,299 `<repo>_commits.json` files (the AI-authored commit lists). **Input to reproduction Tier 3.** |
| `results-out.zip` | 978 MB | ~17 GB | Per-repo analyzer output (`debt_metrics.json`, `issue_survival.json`, `lifecycle_metrics.json`, `destiny_metrics.json`, `summary.json` + `debug/`) for all 6,299 repos. **Input to reproduction Tier 2.** |

`focused_repos.json` and `ai_repos.csv` are also in the GitHub repo; they are mirrored here so this dataset is self-contained. The two `.zip` bundles are **only** available here.

---

## How to download

These files are tracked with Git LFS — use the `huggingface_hub` client so downloads
are counted and resume automatically if interrupted:

```bash
pip install -U huggingface_hub

# A single bundle
hf download yueyuel/tech-debt-ai-coding results-out.zip --repo-type dataset --local-dir .
hf download yueyuel/tech-debt-ai-coding commits.zip     --repo-type dataset --local-dir .

# Or the entire dataset
hf download yueyuel/tech-debt-ai-coding --repo-type dataset --local-dir ./tech-debt-data
```

In Python:

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download("yueyuel/tech-debt-ai-coding", "results-out.zip", repo_type="dataset")
```

---

## Reproduction tiers

The study reproduces at three levels of effort. Full step-by-step commands are in the
[GitHub README](https://github.com/yueyueL/tech-debt-ai-coding#reproduction-tiers).

| Tier | What you do | Time | Disk | Needs |
|---|---|---|---|---|
| **T1** | Browse `results/out/aggregate_summary.json` via the dashboard | <1 min | 0 GB | Python 3.10+ (GitHub clone only — no download) |
| **T2** | Re-aggregate from saved per-repo metrics | ~5 min | ~10 GB | Python 3.10+, `results-out.zip` |
| **T3** | Re-run the full pipeline from raw commits | days | ~500 GB | Python 3.10+, Pylint, ESLint, Semgrep, git, `commits.zip` |

**T2 — re-aggregate:**
```bash
hf download yueyuel/tech-debt-ai-coding results-out.zip --repo-type dataset --local-dir .
unzip results-out.zip -d results/
python3 -m src.reporting.aggregate --out-dir results/out
```

**T3 — full re-run:**
```bash
hf download yueyuel/tech-debt-ai-coding commits.zip --repo-type dataset --local-dir data/
unzip data/commits.zip -d data/
# then follow the GitHub README for analyzer setup + batch run
```

---

## Schema

Brief notes below; full schemas (including the commit and metric JSON shapes) are in
[`data/README.md`](https://github.com/yueyueL/tech-debt-ai-coding/blob/main/data/README.md) on GitHub.

**`focused_repos.json`**
```json
{
  "total_repos": 6299,
  "repos": [
    {"repo": "owner/name", "file": "data/commits/owner_name_commits.json",
     "stars": 12345, "language": "Python", "ai_commits": 42,
     "total_commits": 5000, "ai_percentage": 0.84}
  ]
}
```

**`commits/<owner>_<repo>_commits.json`** (inside `commits.zip`)
```json
{
  "repo": "owner/name", "total_commits_scanned": 5000, "ai_commits_count": 42,
  "tools_found": {"copilot": 30, "claude": 12},
  "ai_commits": [
    {"sha": "abc123", "ai_tool": "copilot", "detection_method": "coauthor",
     "date": "2025-06-15T10:30:00Z", "url": "https://github.com/owner/name/commit/abc123"}
  ]
}
```

---

## License

Released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). You are free
to share and adapt the data with attribution to the paper below.

## Citation

```bibtex
@article{liu2026techdebt,
  title         = {Debt Behind the AI Boom: A Large-Scale Empirical Study of
                   AI-Generated Code in the Wild},
  author        = {Liu, Yue and Widyasari, Ratnadira and Zhao, Yanjie and
                   Irsan, Ivana Clairine and Chen, Junkai and Lo, David},
  year          = {2026},
  eprint        = {2603.28592},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE}
}
```