AI sounds equally sure of itself when it's right and when it's wrong.
That's the part that catches people. A made-up statistic reads exactly like a real one. A wrong date sits in a clean, well-formatted paragraph. A summary quietly drops the one detail that mattered.
The answer isn't to stop using AI. It's to check its work the way you'd check a new hire's: quickly, in the places where mistakes actually happen.
Here's a routine that takes about five minutes and catches most of them.
Why AI gets things wrong
You don't need the technical details. You need one idea: AI predicts what a good answer looks like. Most of the time, what looks right is right. When it doesn't know something, it doesn't stop. It fills the gap with something that looks right.
That's why mistakes cluster in predictable places:
- Specifics: names, numbers, dates, prices, quotes, links
- Recent or niche facts: anything new, local, or obscure
- Math: especially multi-step calculations done "in its head"
- Your context: details about your business it was never told
General explanations and structure are usually solid. Specifics are where you look first.
Step 1: Highlight every checkable claim (1 minute)
Skim the output and mark anything that could be true or false:
- Numbers and percentages
- Names of people, companies, tools, or laws
- Dates and timelines
- Quotes and "studies show" statements
- Links and citations
If a sentence would embarrass you if it were wrong, it's on the list.
You can have AI do the first pass for you:
List every factual claim in your last answer that someone could
verify: numbers, names, dates, quotes, and sources. For each one,
tell me how confident you are and why.
This won't catch everything, but it surfaces the shaky spots fast. Pay attention to anything it calls "likely" or "typically." Those are guesses.
Step 2: Verify the claims that matter (2 minutes)
You don't need to check everything. Check what would cause damage if it were wrong: anything going to a client, into a public post, or into a decision with money attached.
For each one:
- Numbers and facts: find the original source, not a page that repeats it
- Links: click them. Made-up URLs are common, and they often look real.
- Quotes: search the exact wording. If you can't find it, don't use it.
- Math: ask the AI to show each step, or check it in a spreadsheet
If you're using Claude or ChatGPT with web search on, ask it to cite where each fact came from. Then open the source and confirm it actually says that. Sources get misread too.
Step 3: Ask it to argue against itself (1 minute)
This is the most useful habit on the list. Before you use an answer, ask:
Review your answer as a skeptical expert. What's wrong, missing,
or oversimplified? What would someone who disagrees say?
AI is often better at spotting problems in an answer than at avoiding them on the first try. This one prompt regularly turns up a missing exception, a weak assumption, or a step that doesn't hold.
Step 4: Check it against what you know (1 minute)
Read it one more time with one question in mind: does this match my actual situation?
AI writes for the average business, the average customer, the average case. You know things it doesn't:
- Your pricing, policies, and promises to clients
- What already happened with this customer or project
- Rules in your industry or location
- What your audience already knows
A proposal that's correct in general can still be wrong for this client. That last read is yours.
What the routine catches
Say you ask AI to draft a short email to a client about switching their site to new hosting. The draft looks great. The routine finds:
- Step 1: it says the move takes "24–48 hours." That's a checkable claim.
- Step 2: that number is the generic DNS answer. Your actual cutover plan is two hours, scheduled overnight.
- Step 3: asked to critique itself, it points out the email never says what the client needs to do (nothing) or what happens if something breaks (you roll back).
- Step 4: you remember this client was burned by a bad migration last year. The email should lead with how you'll prevent that.
Five minutes, and the email goes from "fine" to accurate and reassuring.
Make checking automatic
Once you do this a few times, build it into how you ask in the first place:
- Give it your facts up front. Paste in the real numbers, policies, or notes. AI makes far fewer things up when it doesn't have to guess.
- Tell it to admit gaps. Add: "If you're not sure about something, say so instead of guessing."
- Save the check prompts. In Claude, put them in a Project's instructions so you don't have to retype them.
This is the same move as building a voice profile: you write the instructions once and reuse them, so the quality is consistent every time.
When to skip the routine
Not everything needs checking. Brainstorming, outlines, first drafts you'll rewrite anyway, and questions where you'll immediately see if the answer works: use those freely.
The routine is for anything that leaves your hands: client emails, published content, numbers in a report, and decisions you can't easily undo.
Where this fits
Knowing where AI fails is the foundation for everything else. You can't build a reliable AI workflow on output you don't trust. The good news: once you know where mistakes happen, checking takes minutes, not hours.
Want the guided version?
How AI works, where it fails, and how to check it is Phase 1 of Monarc University. From there you build prompt systems, connect them into real workflows, and ship a capstone project. The founding cohort opens in early 2027.