Hugging Face's logo

Datasets:
whyamas
/
ai-commit-census-reverse-engineering

Size:
n<1K
Tags:
reverse-engineering
ai
git
commits
decompilation
License:
Dataset card Data Studio Files
xet
 Community
ai-commit-census-reverse-engineering / README.md
whyamas's picture
whyamas
Upload folder using huggingface_hub
6c73e24 verified
preview code
|
Raw History Blame Contribute Delete
5.87 kB
metadata 
license: cc-by-4.0
pretty_name: AI commit census for reverse engineering projects
tags:
  - reverse-engineering
  - ai
  - git
  - commits
  - decompilation
size_categories:
  - n<1K

AI commit census for reverse engineering projects

Version 2026-11-10. Measured on 10 Nov 2026 (snapshot captured 2026-11-10T08:52:25Z).

This dataset counts how many commits in public reverse engineering projects credit an AI tool as co-author or author. At the snapshot, 190,658 of 450,474 commits (42%) across 838 public projects credited an AI. Claude is credited on 183,907 commits and GPT / Codex on 10,022. The census covers every kind of project, games included.

The census is published on the web at https://reversedwithai.com/ai-commit-census, and this dataset is the same data as a file. I run it, and I rerun it about once a week. Each listed project has its own page at https://reversedwithai.com/p/, with the same figures.

How it was measured

  • I count the AI co-author credits that coding tools write into each project's public git history. A commit counts only if it names an AI as co-author or author, typically through a Co-authored-by trailer.
  • The repositories are the 859 that my discovery had found on GitHub by 10 Nov 2026: 508 listed in the directory and the rest awaiting review.
  • Each commit counts once. Commits are matched by hash. Forks repeat their parent's history, so forks are grouped into one project by their first commit.
  • A commit crediting two models counts once for each model.
  • Figures for individual projects are published only for the 508 listed entries. The corpus totals (838 projects) include the unlisted projects, which have no per-project rows in this dataset.

Limits

These limits are copied from the census page.

It is a minimum. A commit counts only if it names an AI as co-author or author, and many tools and people leave that out.

Commit counts measure activity. An agent that commits every few minutes counts for far more than a person who squashes a week of work, and the 10 busiest projects account for 39% of AI-credited commits.

It doesn't show whether the result works. For that, see each project's reported outcome.

Each commit counts once. Forks repeat their parent's history, so commits are matched by hash and forks are grouped into one project by their first commit.

The repositories are the 859 that our discovery had found on GitHub by 10 Nov 2026: 508 listed in the directory and the rest awaiting review. Recently active repositories are easier to find, so recent weeks can look busier than they were. Figures for individual projects are published only for listed entries.

The page footer adds: "Figures are reported by project authors and not measured by this site."

Notes from checking the files against each other (these are not on the page):

  • The weekly series starts on 18 May 2026. Its aiCommits sum to 175,941, which is 14,717 short of the corpus total of 190,658. The JSON does not explain the gap. Some listed projects have a first AI commit before 18 May, so the gap is probably older commits, but I have not confirmed that.
  • The last week row (starting 9 Nov 2026) is partial, because the snapshot was taken on 10 Nov.
  • projects.csv covers only the 508 listed projects: 292,189 commits and 135,180 AI-credited commits. The remainder of the corpus (158,285 commits and 55,478 AI-credited commits, derived by subtraction) belongs to the unlisted projects, whose per-project figures are not published.
  • In families.csv and models.csv the rows overlap, because one commit can credit two families or two models. Their sums (197,293 and 207,000) are therefore larger than the 190,658 AI-credited commits.
  • For 29 listed projects the model rows add up to fewer commits than the project's aiCommits. The JSON does not say why.
  • For 15 listed projects two models tie for the most commits. top_model takes the first one in the order used in data.json.

Files

File Contents
data.json The full snapshot as published by the census. Keys: version, capturedAt, corpus, families, models, weeks, projects (an object keyed by slug), license, attribution.
models.csv One row per model name as credited (15 rows). Columns: model, commits (commits crediting that model), projects (projects crediting it).
families.csv One row per family (8 rows). Columns: family, commits, projects.
weeks.csv One row per week, 26 rows from 18 May 2026 to the week of 9 Nov 2026. Columns: week (the Monday, YYYY-MM-DD), aiCommits (AI-credited commits that week), activeProjects (projects with AI-credited commits that week; the page does not define this, so this is my reading), newProjects (projects whose first AI-credited commit falls in that week, the page's "First AI commit that week" column).
projects.csv One row per listed project (508 rows). Columns: slug, commits (all commits in the project after de-duplication), aiCommits (AI-credited commits), aiDays (days with an AI-credited commit), firstAi and lastAi (first and last AI-credited commit dates, YYYY-MM-DD, as given in data.json), top_model (the model with the most commits in that project; blank when no model is credited). 83 listed projects have no AI credit and so have a blank top_model.

The CSV files are UTF-8 with a header row and LF line endings. data.json is the source, and the CSVs are flattened from it with a short Python script.

Citation

Reversed with AI, "AI commit census for reverse engineering projects", measured 10 Nov 2026, https://reversedwithai.com/ai-commit-census

Licence

CC BY 4.0. You may reuse the data if you credit Reversed with AI and link to https://reversedwithai.com/ai-commit-census.

Link to the census: https://reversedwithai.com/ai-commit-census