How much of F-Droid is AI generated?

A manual audit of 102 popular F-Droid apps published September 15, 2026 judged 72.5% (74 apps) as mostly AI-written, 9.8% as ambiguous, and only 17.6% (18 apps) as showing little to no signs of AI involvement. The audit used repo heuristics — commit patterns, README style, agentic infrastructure — not code-quality analysis.

How reliable is the F-Droid AI audit methodology?

The author states there is no reliable way to detect LLM-authored code from text alone, so the ratings are based on aesthetic and commit-history heuristics, with any repo using agentic tooling automatically classed as mostly AI. It is an informed visual audit, not a verified measurement — which is itself the point about the missing provenance layer.

What does GitClear's report say about AI-written code?

GitClear's Maintainability Gap report, covering 623 million analyzed code changes from 2023 to 2026, found block-level code duplication rose 81% over 2023 levels — a corporate-side measurement that, landing the same week as the F-Droid audit, points the same direction from the opposite end of the software ecosystem.

F-Droid: 72.5% of Popular Apps Mostly AI-Generated

What shipped

Today a student developer going by tintotint published a manual audit of 102 popular F-Droid apps, scored by repo aesthetics — commit patterns, README smells, agentic infrastructure. The tally: 74 apps (72.5%) judged "mostly AI" (expected >50% LLM-authored code), 10 (9.8%) genuinely ambiguous, and just 18 (17.6%) showing little or no signs of AI involvement. Anything with a coding harness wired to the repo auto-landed in the top tier. The post hit the Hacker News front page within hours. The author's own caveat is the part everyone should read twice: there is no slop detector, no provenance metadata, no way to verify any of it — the ratings are trained eyes on recent commits.

The connection nobody made

Yesterday GitClear's Maintainability Gap report put corporate numbers on the same tide: 623 million analyzed changes, block duplication up 81% over 2023. One dataset is measured across enterprise monorepos; the other is eyeballed across the hobbyist app shelf. Different methods, opposite ends of the software economy, same conclusion landing within 24 hours of each other. But the sharper point is the channel, not the code. F-Droid exists because people don't trust app stores — curation and human provenance are its entire value proposition. Nobody browses F-Droid for convenience. If the anti-store is now 72.5% machine-written by the audit's own heuristic, the trust differentiation has quietly inverted, and there's no metadata anywhere to prove otherwise.

Verdict

Don't read this as a quality verdict — the author explicitly skipped code analysis, and some of the 72.5% have real users. Read it as a provenance measurement. Human-written FOSS still exists (17.6% here), but it's no longer the default anywhere, including the shelf built to protect it. The ask that follows is boring and urgent: repos and registries should ship signed, machine-readable authorship metadata so the next audit isn't vibes. Until then, "is this slop?" stays a question only commit archaeology can answer.

What shipped

The connection nobody made

Verdict

Frequently Asked Questions

Related reads