How many AI-generated commits are on GitHub?

As of February 2026, over 5.3 million AI-attributed commits per month are detected on public GitHub repos, spanning Claude Code, Jules, Devin, Aider, Gemini Code Assist, and OpenAI Codex CLI. The true number is higher as Copilot, Cursor, and Gemini CLI leave no commit-level markers.

How fast is AI-generated code growing on GitHub?

Claude Code grew from 24 commits in January 2025 to 5.19 million in February 2026 — a 216,000x increase in 13 months. The best-fit model is exponential with a doubling time of approximately 1.38 months.

What percentage of GitHub code is written by AI?

AI-attributed commits represent about 7.9% of all public GitHub push events. Adjusting for larger diff sizes in AI commits (2.5x), the estimated code volume share is approximately 19.6%. This excludes Copilot and Cursor which leave no markers.

How does botcommits.dev detect AI commits?

Detection relies on commit attribution markers: Co-Authored-By trailers (e.g., Claude @anthropic.com), author emails (noreply@anthropic.com), bot actors (google-labs-jules[bot]), and commit message patterns (e.g., 'Generated with [Claude Code]'). Data comes from GH Archive via BigQuery (Jan-Oct 2025) and GitHub Search API (Nov 2025+).

botcommits.dev - Tracking AI-Generated Commits on GitHub

AI-Attributed Commits

Repos with AI Commits

Developers Using AI

AI Share of Pushes

Growth Trajectory: Exponential vs Logistic vs Gompertz (+ 6-month projection, shaded = private repo extrapolation)

Exponential y = a·e^bt — R²=0.989, AIC=314.7

Logistic y = L/(1+e^-k(t-t₀)) — R²=0.989, AIC=316.7, ceiling hit upper bound

Gompertz y = L·e^{-e^-k(t-t₀)} — R²=0.981, AIC=323.8, ceiling ~641M

Best fit by AIC: Exponential. The logistic ceiling converged to the upper bound (~1B), meaning the data does not yet show saturation. With only 14 months of observations and a data quality gap at month 10, we cannot distinguish exponential from early-stage logistic. This does not predict indefinite exponential growth — it means the inflection point has not yet been observed.

Shaded band: Upper bound estimates total commits (public + private repos). 81.5% of GitHub contributions happen in private repos (GitHub Octoverse 2025), so the upper bound = public data ÷ 0.185. This assumes equal AI adoption in private repos — a speculative extrapolation, not measured data.

The Iceberg: Public Repos Are Only 18.5% of GitHub Activity

~28M

Estimated total Claude commits/month

(public + private extrapolation)

81.5%

of GitHub contributions are in private repos

Source: GitHub Octoverse 2025

Copilot + Cursor commits (unattributed)

No attribution markers — invisible to this method (+ Gemini CLI)

Methodology: We observe 5.19M Claude commits in Feb 2026 from public repos. Since 81.5% of GitHub activity is private, and assuming similar AI adoption in private repos, the total is estimated at 5.19M ÷ 0.185 ≈ 28M commits/month. This is speculative — enterprise adoption may be higher or lower than public repo rates. Copilot, Cursor, and Gemini CLI leave no commit-level attribution markers and are excluded entirely.

Commits by AI Tool

AI vs Human: Estimated Code Volume (commits + diff size adjustment)

Denominator: total push events from BigQuery (not total commits — each push averages ~3 commits). Diff size adjustment: AI commits have ~2.5x larger diffs than human commits (per published research), so the code volume share is estimated at 2.5x the raw commit share. Both lines shown. Copilot/Cursor/Gemini CLI are excluded (no attribution markers) — real AI share is higher.

Growth Rate (MoM %)

Acceleration (Change in Growth Rate)

Cumulative AI-Attributed Commits (Jan 2025 - Mar 2026)

Tool	Total Commits	Peak Month Repos	Peak Month Devs

Methodology & Data Notes

Data sources

Jan - Oct 2025: GH Archive via Google BigQuery. Exact commit-level counts from PushEvent payloads.
Nov 2025 - Feb 2026: GitHub Search API (approximate). GH Archive stopped including commit message payloads in Nov 2025, reducing table size from ~400 GB/month to ~89 GB. We fall back to GitHub's commit search endpoint, which returns total_count estimates that may differ from exact counts.
Mar 2026*: Forecast. 1,723,910 commits observed in 8 days (Mar 1-8), linearly extrapolated to 31 days.

Detection signals

Claude Code: Co-Authored-By: Claude ... @anthropic.com trailers, Generated with [Claude Code] body markers, noreply@anthropic.com author email.
Aider: aider: prefix in commit messages, Co-Authored-By: aider trailers.
OpenAI Codex CLI: Co-Authored-By: ... @openai.com, noreply@openai.com author email.
Jules (Google): google-labs-jules[bot] actor/author. Commits authored directly by the Jules GitHub App bot account.
Gemini Code Assist: Co-Authored-By: gemini-code-assist[bot] trailers. Only captures applied code review suggestions, not all Gemini-assisted code.
Devin: Jan-Oct: devin-ai-integration actor/author in BigQuery. Nov+: co-authored-by devin-ai via GitHub Search API (author-based search returned 0; co-author trailer search works). Detection method differs between periods.

Known limitations

Oct 2025: GH Archive data appears incomplete for this month (61M push events vs ~71M in adjacent months). Commits counts are artificially low. This month is excluded from growth rate and acceleration calculations.
Nov 2025 growth rate: Since Oct is excluded, Nov's MoM growth is estimated as the average monthly growth from Sep to Nov (two-month span halved).
Growth rate charts start at Apr 2025: The first 3 months (Jan-Mar) have extreme MoM percentages (e.g., 8,800% Jan to Feb) due to near-zero base counts. These are excluded from the growth rate and acceleration charts to keep the scale readable.
GitHub Copilot, Cursor, and Gemini CLI do not leave commit-level attribution markers, so they are invisible to this methodology. The numbers here represent a lower bound of AI-assisted coding.
Only public repos are tracked. Private repo AI usage is not measurable via GH Archive or the Search API.

Model fitting

Three models fitted to Claude commit counts using scipy.optimize.curve_fit (nonlinear least squares): Exponential (2 params), Logistic (3 params), Gompertz (3 params).
Oct 2025 excluded from fitting due to incomplete data. 13 data points used.
Model comparison via AIC (Akaike Information Criterion): lower = better fit with parsimony penalty for extra parameters.
Exponential wins (AIC=314.7 vs 316.7 logistic, 323.8 Gompertz). The logistic ceiling parameter converged to its upper bound (~1B), meaning the optimizer found no evidence of saturation in the data. Effectively, the logistic degenerates to an exponential with an extra unused parameter.
This does not predict indefinite exponential growth. It means the inflection point of the S-curve has not yet been observed with 14 months of data. More data points are needed to distinguish the models.
Projections beyond Feb 2026 are extrapolations and should be treated with appropriate skepticism. The exponential projects 104M commits/month by Aug 2026; the Gompertz projects 40M. The true value likely falls between them.

Open source

Raw data (CSV), BigQuery queries, Python analysis scripts, and this dashboard are all available at github.com/edgardomunoznajar/github_explosion. Contributions and corrections welcome via pull requests.