# What public commits can tell us about AI assistance

Canonical: https://devimprint.com/articles/how-ai-coding-tools-sign-commits/

Published: 2026-09-04T00:00:00.000Z
Author: DevImprint
Updated: 2026-09-13T00:00:00.000Z

How to interpret observations of AI-assisted public development, including attribution uncertainty, coverage and the limits of comparison.

Public repository history can contain evidence of AI assistance. DevImprint
associates that evidence with coding tools and reports observations within a
defined time window. It does not inspect private work or verify who wrote
individual lines of code.

## What an observation means

An observed association is evidence of public attribution. It does not establish
how much of the work a tool performed, whether the code is correct, or whether
the attribution is accurate. A small change and a large refactor each count as
one commit.

False positives and missed activity are possible. Counts are observations under
this method, not verified totals or guaranteed lower bounds on AI use.

## What we publish

The public methodology describes the population, ranking definitions, UTC time
window, scan coverage and known limitations. Downloads and citations carry the
source snapshot timestamp and digest so a reader can identify the measurement.

Exact matching rules, attribution identifiers and detection examples are not
published, to limit manipulation. The [tool pages](/tools/) provide observed
activity and its scope. See [the methodology](/methodology/) for the public
measurement contract and correction process.

## Why absence says little

Private work is outside the measurement. Repository coverage is incomplete, and
attribution practices differ across tools, teams and workflows. Some assistance
leaves no observable public evidence. A developer can use AI extensively without
appearing in these observations.

Changes in observed volume can reflect changes in coverage or attribution as
well as changes in activity. Tool counts are not user totals or market share.

## How to use the figures

Read the population and time window beside each statistic. The published ranking
is a bounded subset; its rows do not sum to the complete population totals.
Separate people from automation, and compare observations only with their
coverage and attribution limitations in view.

Use the citation control to include the value, scope, source date and coverage.
A profile citation records an observation at that time; the snapshot archive
preserves aggregate context, while personal records remain subject to correction
and removal.

If an association looks wrong, use [the correction process](/methodology/#corrections).
To remove a personal record, use [the removal request page](/data-request/).