AI Code Works Perfectly Until the Requirement It Wasn't Told About Arrives

AI-generated code fails at the exact moment it succeeds too well: it solves the requirement in front of it so precisely that every assumption gets frozen into the structure, and nothing is left loose enough to bend later.

The Details:

• Ask an AI to process a payment and you get working code where the currency is hardcoded, the provider's SDK is called directly, and tax logic assumes one jurisdiction. Each choice is correct for the question asked. Eighteen months later, expanding to a new region turns what should be a config change into a multi-week rewrite because those assumptions are scattered across a dozen files.

• A human architect asks a second question the model never asks on its own: what about this is likely to differ in six months? That question is what turns a hardcoded literal into a parameter and a direct API call into something sitting behind an interface. AI reproduces the shape of what you request, not the shape of what might come next, so the seam only appears if someone names it first.

• The fix costs almost nothing: naming one or two likely axes of change before sending the prompt, like currency, provider, or region, changes the resulting architecture without making the model smarter. The asymmetry favors guessing: a wrong guess costs nothing, a missed one costs a rewrite.

• This spreads through teams via small rituals, not process. A one-line comment naming an assumption, or a reviewer noting that a call was placed behind an interface, does more than any architecture review. Ceremony gets routed around; recognition sticks.

Bottom Line: The gap AI cannot close is the requirement you haven't told it about yet, so closing it stays entirely on you.