Bulldozer of Confidence
Charge first, ask questions never
The Bulldozer doesn't question AI output because questioning things is slow, and slow is the enemy. When the model generates a function, the Bulldozer reads the first three lines, decides it looks right, and moves on. The model makes this worse by mirroring and amplifying that confidence.
- Symptom
- Large, fast pull requests that work on the happy path but crumble under edge cases. Code reviews met with 'it works, though.' When AI-generated code breaks, they don't analyze the failure — they prompt the AI again and replace the broken code with new generated code.
- Why it matters
- The Bulldozer's damage accumulates gradually and then arrives all at once. Individual commits look fine. But the codebase develops a peculiar quality: it works without anyone understanding why it works.
- What the chapter gives you
- How to spot the inverse relationship between commit frequency and review depth, why brute-force debugging with AI is a downward spiral, and the counter-moves that slow the Bulldozer just enough.
Parent Class
From Volume 1 of The AI Developer's Field Guide
Read the full chapter in The AI Developer's Field Guide.
Get the book
Or choose your store:
Recognize this one in your codebase?
The book has the full chapter, the symptoms, and the interventions.