The most dangerous thing about AI isn’t that it’s sometimes wrong. It’s that it is wrong in exactly the same confident tone it uses when it’s right. There’s no stammer to listen for. If you publish AI-assisted content — videos, posts, product pages, client work — you need a checking habit. Here’s mine.
Sort claims by blast radius
Not everything needs checking. “The sky appears blue due to light scattering” — fine. But three categories get verified every time:
- Numbers — prices, percentages, dates, statistics. AI mixes up years and invents plausible figures constantly.
- Anything time-sensitive — model names, tool features, platform rules, what something costs. AI’s knowledge has a cutoff date; the world doesn’t.
- Anything with consequences — health, money, legal, safety. Wrong here doesn’t cost views, it costs trust or worse.
A snippet is not a source
When you search to verify, don’t stop at the search-result preview text — it’s often AI-generated too, or quoting the same wrong article. Open the actual page. Find the original source: the company’s own pricing page, the study itself, the official documentation. The test I use: have I actually seen the primary source, or have I only seen someone describing it?
Ask the AI to cite, then check the citations
Asking “what’s your source?” does two useful things. Sometimes the AI admits it’s unsure — valuable. Sometimes it names a source — which you then open, because AI also invents citations that sound real, right down to plausible dates.
Label what you couldn’t verify
Sometimes you can’t confirm a claim and want to use it anyway. Say so: “reportedly,” “as of early 2026,” “I couldn’t verify this.” Readers forgive honest uncertainty. They don’t forgive confident nonsense.
Make it a checklist, not a virtue
The failure mode is believing you’ll “be careful.” You won’t, at 11pm, on video #40. What works is mechanical: before anything publishes, list the factual claims, mark each ✓ verified / ✗ removed / ⚠ labeled. Five minutes. I’ve had entire scripts survive review untouched, and I’ve had ones where three “facts” died at the checkpoint — those three were the ones that would have been in the comments forever.