GitHub counts an AI project by its code imports, not a survey

The setup

Octoverse is GitHub's annual report on activity across its own hosted repositories, and the 2025 edition, published 28 October 2025, is built entirely from platform data: commits, pull requests, contributor counts and repository metadata, rather than a survey asking developers what they think about AI. The report itself states its counts come from automated classification of public repositories, with a documented methodology explaining how categories like AI project and agentic workflow are assigned rather than self-reported by repository owners.

What the documents show

GitHub's own methodology notes state a repository is classified as an AI project through signals including generative-AI SDK usage detected in its imports and dependencies, plus repositories tagged with topics such as AI, ML or LLM, alongside a broader classification method the report does not fully enumerate. Agentic workflows are defined separately, as development tasks completed with the help of autonomous or semi-autonomous tools, and the report gives Copilot's coding agent, which creates pull requests, triages issues and runs tests, as its example. The report's landing page states more than one million pull requests were created by the coding agent between May and September 2025, a usage count GitHub reports from its own systems, not a developer's self-assessment.

The friction

Because the AI-project classification runs on dependency imports and topic tags, it will catch a repository importing a generative-AI library for a small, experimental use and count it the same as a production system, a broad-net method the report itself describes as capturing general experimentation and adjacent projects rather than a narrower, verified category. The methodology also applies minimum thresholds, such as a minimum contributor or repository count before a ranking is published, to reduce noise, which means some categories are filtered before publication in ways a reader comparing two numbers might not notice unless the methodology is read alongside the headline figures.

What changed in the work

What this year's Octoverse supports is a GitHub-scale count of AI-tagged repository growth and coding-agent pull-request activity in 2025, measured from GitHub's own systems rather than asked of developers after the fact, a different and more concrete kind of evidence than an opinion survey. It does not support a claim about how developers feel about these tools; the classification methodology is entirely about what code and metadata show, and a sentiment claim would need a separate survey instrument, which this edition does not include.

• Is the Octoverse figure being cited a usage count, or is it being treated, incorrectly, as a sentiment finding?
• Would the team's own AI-tagged repositories pass GitHub's dependency-import and topic-tag test, or are they undercounted by it?
• Has the methodology's minimum-threshold filtering been checked before comparing two category counts directly?

GitHub's own methodology section is the part of Octoverse worth reading before quoting its headline numbers: it draws a clear line between what dependency imports and pull-request counts can show about AI-assisted activity, and the developer-opinion questions this report's data was never built to answer.