AI Joe
← Blog· AI Axiom

When Requirements Shift, AI's Tone Doesn't

November 20, 2026

AI coding assistants sound exactly as confident after a requirement changes as they did before it. That steady tone is the problem, because underneath it code quietly stops matching reality while every test still shows green.

The root cause is structural. Fluent explanation used to be a proxy for correctness because only a human who truly understood a system could narrate it clearly. Language models broke that link: generating polished, confident prose costs them nothing, whether or not they traced every downstream effect of a change. A model will patch the function you pointed to and miss three callers that depended on the old rule, and it will describe the fix in the same assured register either way.

The Details:

  • Tests written under an old assumption confirm that assumption, not the new requirement. Shift a data model from one record per user to many, and the old tests still pass because they were built to check the one-record world, not the new one. Green checkmarks become false reassurance rather than evidence.

  • The fix is to name dependencies before touching code. List every caller, index, fixture, and serializer that the old requirement was propping up; a vague list signals the model is pattern-matching rather than understanding the system, while a specific one makes the change nearly mechanical.

  • A cheap diagnostic: ask the model to predict three things that will break before it edits anything. Its answer reveals whether it has a real model of the system or is guessing from the file in front of it, and pushing back on its first answer shows whether it can defend the reasoning or simply rewrites to please you.

  • Keep a lightweight running ledger near the code, short notes like 'this assumes X,' so assumptions get searched later instead of rediscovered under pressure during the next change.

Bottom Line: When requirements move, trust in AI output should reset to zero, because the model's tone will never tell you that its footing just got shakier.

Enjoy this article?

Listen to the Claude Code Conversations radio show or join the community.