When generated code breaks something in production, the person who approved it answers for it, not the tool that wrote it. That fact hasn't changed with faster, more fluent AI coding tools; it has just gotten easier to forget.
Why it matters: As AI-written code ships faster than teams can meaningfully review it, the habits that once forced careful judgment are quietly disappearing, and nobody has updated who is accountable when that judgment goes missing.
The Details:
- AI-generated code can compile, pass tests, and read cleanly while still being the wrong choice for a specific system at a specific moment. The model has no memory of which boundaries failed before, no sense of organizational risk tolerance, and no stake in what happens after deployment.
- Writing down acceptance criteria and non-negotiable failure modes before the first prompt gives reviewers a fixed standard to check output against later. Without that document written in advance, the bar for acceptable quietly slides whenever a deadline gets close.
- Reading a generated diff line by line, rather than skimming it, matters more now because generation feels nearly free. That cheapness makes people reluctant to discard flawed code, even though nothing was actually invested in producing it in the first place.
- Certain surfaces, like authentication, payment handling, or personal data, should require a second human reviewer as a standing rule, set before any specific deadline exists. Deciding this in advance keeps urgency from making the call by default when nobody is watching.
Bottom Line: AI made building software cheaper; it did not make being wrong cheaper, and that gap is where the architect's job actually lives now.
