A decade after a Java‑vs‑engineer debate, the conversation has shifted to AI‑generated code. The author distinguishes a vibe coder, who spins a prototype to test an idea, from a software engineer, who owns the full development lifecycle. AI can now write code faster, but the real question is how that output survives real users, data, compliance and incident response.

Teams that measure success by “time to first working version” reward rapid demos but ignore review, testing and rollback costs. The author proposes “time to safe merge” as the proper metric: it counts reviewability, risk, test quality, ownership and the ability to explain every change. If AI lowers generation cost while raising safe‑merge cost, overall efficiency stalls.

Because models lack system‑wide context—incident history, migration quirks, security policies—engineers must constrain prompts and own the resulting code. A responsible engineer turns AI output into a narrow, explainable change that can be merged without dragging unrelated files. The takeaway: AI is a productivity aid, not a shortcut that bypasses the engineering discipline.

Even open‑source maintainers echo this view; projects that ban AI contributions cite noisy pull requests that break legacy behavior. Rather than forbidding the tool, the author suggests placing AI work behind the discovery‑delivery line, allowing rapid spikes but demanding strict review before production. In practice, the metric shift forces teams to treat AI‑generated snippets as drafts, not final commits.