Trust, But Verify — The Checklist That Keeps AI Honest

I spent thirty years in cybersecurity, where the first rule is that confidence is not evidence. AI models are the most confident-sounding technology ever built — right or wrong, the tone never changes. That's not a reason to avoid them. It's a reason to have a verification habit.

The AI Output Verification Checklist is mine, written down.

What's on it

The checklist walks you through the questions to ask before you act on AI output: Which claims are checkable facts, and did you check the ones that matter? Are the numbers internally consistent? Do citations actually exist and say what they're claimed to say? Would a domain expert wince at anything here? And the one people skip — does this output actually answer the question you asked?

It's calibrated by stakes. A brainstorm doesn't need an audit; a client deliverable, a legal paragraph, or a number in a board deck does. The checklist tells you which level of scrutiny fits which job, so verification stays a habit instead of becoming a burden.

The mindset shift

The goal isn't distrust — it's appropriate trust. Once you have a repeatable way to verify, you stop hedging and start delegating bigger work to AI, because you know your safety net catches what matters. The people getting the most out of these tools aren't the most trusting. They're the best at checking.

This checklist is the heart of chapter 3 of Creating with Claude. The book teaches the habit; this page is the habit.