AI-generated code quietly encodes dozens of small decisions that nobody asked for, and those decisions vanish without warning the moment you request a change.
The Details:
Every time an AI resolves ambiguity in a spec, it makes a choice: which sort order to use, what an error message says, which extra fields to include in a response. None of these choices get labeled or documented, so the code itself becomes the only record of a decision that was never actually requested or reviewed.
Teams start treating that unlabeled behavior as a stable contract. A frontend keys off a specific error string. A dashboard consumes a field nobody explicitly asked for. The dependency grows silently, and the model has no way of knowing it was ever created in the first place.
When a change request comes in, the AI starts from a blank context. It has no memory of why the code looked the way it did, so a load-bearing quirk reads as an arbitrary implementation detail. Tests pass, the diff looks clean, and the behavior people relied on disappears without any alarm going off.
The fix is a single question asked right after code is written, not weeks later: what did you decide that was not in the ticket. The answer surfaces the hidden choices in a few sentences, giving someone a chance to flag the risky one before it ships. Anything that genuinely matters then gets pinned down as a test, turning an accidental behavior into a promise the model can actually see on the next pass.
Bottom Line: Code written by AI carries invisible agreements, and the only way to keep them from breaking is to make them visible before the next edit erases them.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
