Letting AI draft system design before humans sign off — your comfort level?

Options

• Brainstorm only — diagrams and options, never authority to merge without a human architect.
• Comfortable if every claim links to a doc, metric, or test we already trust.
• Not OK for production paths — hallucinated confidence is worse than no answer.
• Fine for greenfield spikes and internal tools with forgiving users.
• Yes when tests, observability, and feature flags make wrong designs cheap to unwind.
• Design review becomes arguing with a fluent intern who never slept.
• We should require adversarial prompts: one model proposes, another critiques.
• Seniors will outsource thinking and juniors will not learn why boundaries exist.
• Legal and security teams will block this regardless of what developers prefer.
• The win is speed to a first draft, not speed to a correct draft — know the difference.