The problem is not the AI code, but nobody knows anything anymore

The piece argues that the real issue is not the code generated by AI but the erosion of human understanding, intent, and architectural knowledge within teams. In many fast-moving companies, management pushes constant shipping while engineers outsource specs, tickets, tests and even bug triage to large language models, leaving nobody who actually understands why decisions were made or how systems fit together. AI can produce competent, average code and can raise a poor codebase up to that level, but it does not replace domain knowledge. Data engineering historically demanded comprehensive product and business understanding, and while AI removes some friction, it also makes that deep expertise appear unnecessary. Product managers can now spin up viable prototypes without coding skills, but that accelerates poor foundational choices when fundamentals and system design are absent.

The concrete consequence is mounting maintenance burden: rapidly generated pipelines, apps, or dashboards become harder to sustain when no one knows their intent or architecture. AI cannot self-direct; humans remain necessary to provide intent, taste, design and orchestration. The symptomatic cure is retaining and cultivating engineering judgment - hiring and training people who understand systems, preserving documentation and architectural intent, and resisting a culture that values raw output over comprehension and long-term maintainability.