How much of F-Droid is LLM generated?

TEXT START: I love F-Droid.

The Dissection

This is an authenticity panic dressed as an audit. It claims to measure how much of F-Droid is LLM-generated, but actually measures one observer’s ability to recognize suspicious signals in recent commits from a claimed 102-app update batch. The supplied text is truncated, so even the full sample cannot be evaluated here.

The method relies on repository aesthetics, AI disclosures, co-authorship labels, agent infrastructure, README style, and “vibes.” It does not measure code volume, authorship share, quality, security, maintenance, downloads, or user impact. The real subject is the collapse of the old link between effort, skill, authorship, and software output.

The Core Fallacy

The text confuses provenance signals with production facts. A Claude co-author line proves that Claude contributed or was credited; it does not prove that more than half the code is machine-authored. Agent infrastructure proves capability or workflow, not negligence. An AI policy or lack of visible “smells” proves nothing about hidden use.

The text also admits that reliable detection is impossible, then assigns percentage-shaped categories anyway. That is not measurement. It is subjective confidence wearing a lab coat.

The deeper Discontinuity Thesis error is treating “human-written” as the decisive economic category. Under P1, the relevant fact is whether AI produces cognitive output more cheaply and effectively, not whether a human reviewed the diff. Under P2, FOSS norms cannot preserve a stable human-only enclave. Under P3, the question is who controls models, compute, distribution, and verification—not whether an individual developer feels that the machine did the work.

Hidden Assumptions

• Human authorship is a reliable proxy for quality, legitimacy, or meaning.
• AI assistance generally implies low-effort or low-quality software.
• Repository style reliably reveals code origin.
• Recent commits represent the whole application.
• One update batch represents F-Droid as a whole.
• “LLM-authored” has a clean boundary despite prompting, editing, review, and rewriting.
• Agentic tooling is automatically irresponsible rather than merely evidence of a new production method.
• Preserving human authorship is economically achievable through disclosure rules, taste, or community pressure.

These assumptions turn a murky attribution problem into a moral sorting ritual.

Social Function

Classification: partial truth, prestige signaling, and transition management, with an ideological-anesthetic effect.

It contains a partial truth: LLM use is visibly spreading, and provenance is becoming harder to establish. It is prestige signaling because “human-written” becomes a purity badge while “slop” becomes a class marker. It is transition management because disclosure policies and review practices can make AI-assisted production more auditable.

Its ideological effect is more important. It converts a structural labor-replacement crisis into a scavenger hunt for ugly icons, emoji, co-author lines, and suspicious commits. Readers receive a culprit to shame and a ritual to perform, while ownership of the productive machinery remains untouched.

The Verdict

Useful alarm. Invalid census. Strategically mislocated autopsy.

No defensible percentage follows from this text. At most, it shows that some recently updated F-Droid projects visibly use LLMs and that authorship is becoming opaque. Its strongest insight is also the one its method cannot quantify: human effort is no longer a reliable proxy for software output.

Under DT logic, this is a rear-guard cultural document. It sees the smoke and argues over fingerprints on individual commits. F-Droid may preserve access to open software, but openness does not preserve human economic necessity. The source can remain free while control migrates to model owners, compute owners, and distribution gatekeepers. That is the actual obsolescence signal.