My team tracks how much of our codebase is AI-generated. The number shocked us.
We deployed Buildermark last week. It's an open-source tool that scans Git history and flags AI-written lines.
Why We Started Measuring
Every startup has that moment.
You're reviewing a PR and realize you can't tell who wrote it. The human or the AI.
We hit 40% AI-generated code by volume. Some files were 90%.
The CTO asked for the report. Then asked what it meant.
Nobody had an answer.
The Three Problems Nobody Talks About
→ Problem 1: Ownership blur
When AI writes the fix, who owns the bug?
We found junior devs treating Claude output as gospel. They'd copy-paste without understanding.
Senior engineers would approve because "it looks fine."
→ Problem 2: The review gap
Human-written code gets scrutinized. AI-written code gets rubber-stamped.
We caught security issues in AI-generated config files. Stuff a human would never write.
→ Problem 3: The bus factor
If your AI provider degrades (like Claude did last month), your velocity tanks overnight.
We're now vendor-locked to Codeium's style. Claude's patterns. GitHub Copilot's idioms.
What We Changed This Week
We added a pre‑commit hook that tags AI‑generated lines.
Every PR shows the percentage in the description.
If it's over 50%, it needs extra review. No shortcuts.
We also started tracking "AI debt" – lines that only one person understands because they came from a prompt nobody wrote down.
The Real Metric That Matters
Lines of AI code is vanity.
The real metric is: How many AI‑generated lines survive to production without a human understanding them?
We're at 12%.
That's 12% of our codebase that could break and nobody would know why.
Is your team measuring AI code?
What percentage would surprise you?
👇
Top comments (0)