AI-assisted coding is a force multiplier when you already understand the system you're building, and a debt generator when you don't. A balanced look at where it accelerates real work — boilerplate, glue code, refactors with a clear specification — and where it quietly fails: concurrency, performance budgets, and architectural decisions that need a model of the whole system, not the current file.

Most commercial proctoring tools solve the wrong problem: they ship a kernel-level agent, mine biometric data to a third party, and still produce evidence instructors can't audit. A privacy-respecting alternative runs in the browser — face presence via face-api.js, tab-switch and clipboard interceptors, full-screen enforcement — and streams typed events over a JWT-authenticated WebSocket (`/ws/proctor`) into a `ProctoringEvent` collection linked from each `Result.proctoringLog`, so every signal becomes auditable evidence rather than a verdict. Self-hostable, MIT-licensed, and good enough for real classrooms.

Detection accuracy on closed benchmarks is no longer the bottleneck — generators move faster than detectors can be retrained, and a single-modality classifier is a brittle defense. The real fix is system-level: provenance signals at capture, multi-modal verification at inference, and human review pipelines that treat detector output as evidence, not a verdict.