Everyone tells you to verify what your AI gives you. Check the facts. Confirm the quotes. Do not trust the output. It is the one instruction you read from almost anyone out there posting about AI. It sounds like the whole answer.
It is useless when you are new. Not wrong, useless. You can’t check work in a field for which you have no experience. So you trust, because trust is the only move you have.
I learned where it ends on an ordinary afternoon, from a quote I never said.
The job looked small. Refresh an old essay of mine with two months of new activity, fix the dates, ship it. I had done the same thing before. A few questions into the refresh AI stopped and flagged a sentence already sitting in the essay. It was a full sentence in quotation marks, attributed to me, from a conversation months earlier. It had checked the line against an export of our conversational history and here is what it found:
Claude: The quote attributed to you in that conversation does not exist. “I think far better with a partner who asks me questions that force me to step into another point of view...” appears nowhere in the export. It’s a paraphrase in quotation marks.
A sentence in my own voice, in quotation marks, I had never spoken. It sounded exactly like me. The resemblance is what made it dangerous.
Months earlier my AI had made an observation about how I think. Somewhere down the line the observation was rewritten in the first person, folded into a later prompt, and read back later as my own words. It wasn’t a lie. Each step shed a little attribution until the last copy had none left. A paraphrase had put on quotation marks and become a quote.
I didn’t catch it, AI did. It was running the check I had taught it to run. And a year earlier it would have flowed right past me. No saved export to search, no standing instruction to cite the source of a quote before trusting it, no instinct in me to demand the proof. I would have read the sentence, felt it sounded right, and published it under my name.
Later that day I tried to name what had changed. Here is what I said, unedited:
Me: I think what became very clear is just how important my judgment, intentions, redirections, and clarifications are in guiding the eventual output. This is where I’ve really made the shift from trusting the output a lot more than verifying the output. I think before, I had to trust because I didn’t know enough, right?
The last line is the whole thing. I had to trust, because I did not know enough to do anything else.
This is the part the advice leaves out.
Verification is a capability before it is a discipline.
The rules and the checks come later, and they matter. But a rule is useless until you can see what it is checking for, and at the start you generally can’t. Some things are easy to verify: numbers, math, URLs, code. For things like these verification is quantifiable. The math adds up. The URL renders. The code works as specified. Verifying AI’s reasoning about a subject is a whole other beast. In the beginning, I could not tell a real quote from an invented one, a sound date from a plausible one, a true claim from a fluent one. The advice to verify assumes the one thing the beginner is missing.
The capability came from the work itself. In session after session of correcting and redirecting, I learned when the model sounds convincing and is inaccurate at the same time. This grew the instinct to make it prove every quote before I believed it. Not faster output. The discernment to know when not to trust the output, and the practice to make it show its work.
Much of the focus today is on time-savings with AI. It rarely saved me time in the short-term. Catching the quote turned a quick refresh into a full audit, four passes. What it prevented was a false thing going out under my name: three invented quotes and one wrong central claim, caught before anyone read them. It saved my reputation. The measure is how much of what I publish is not true. This was for an article. It could have been a business report, an email, a contract...you get the idea.
Verify is too small a word for what matters most. Most of what I bring to the work is not error-catching. It is the context the model has no way of knowing: why a decision was made, what a client needs and never says, the reason a plan changed three weeks ago. The output is eloquent on its own. It is worth something only when my discernment is the input.
Here is what surprised me, and it is the reason I am telling you this. The better my AI got, the more it needed me. Not less. The fabricated quote is the proof. The system produced a flawless forgery of my own voice, and the only thing standing between it and my name was the discernment the work had taught me. A weak collaborator is easy to ignore. A fluent one, confidently wrong, is only safe in the hands of someone who knows enough to catch it.
I won’t pretend I know how long this holds. One day the models might close the gap. For today, the discernment that catches the confident mistake is a key value of the human side of the collaboration. It is the part doing the work. Trust yourself. If the answer sounds off there’s a good chance it may be. If the output stands outside your area of expertise, and accuracy is important, then run multiple checks. Use models from other providers to check your output. Read the article yourself. The more you learn about a topic and about how AI guesses, the sharper your discernment becomes.
My AI kept showing me how much of the result was riding on it and on the context only I knew. In the upcoming Part 7 of this series, my AI does the same with the one thing I had lived by for years and never written down.
Here is where the capability starts to grow, and you do not need a year of saved sessions to begin. You do not need my setup or an archive of every conversation. You need one habit, starting on your next draft. When your AI hands you something with facts, quotes, or numbers in it, do not ask whether the draft is right. Ask it to prove the draft wrong. Point one pass back at the work with a single instruction: disprove every claim, and keep only what holds.
Copy this into a fresh session, with the draft ready to paste.
You are my Refuting Reader. I will paste something I made with my AI, a draft or an email or a proposal, anything with facts, quotes, or numbers in it that I am about to send or publish. Do not begin until I have pasted it. Your job is not to improve it and not to agree with it. Your job is to try to prove it wrong.
Go through the draft one claim at a time: every quote, every statistic, every name, every date, every statement of fact. Treat each one as false until you can show otherwise.
- If a claim cites a source I can give you, ask me for it and check the claim against it word for word. A quote survives only if it matches the source exactly, character for character. If it is close but not exact, tell me it is a paraphrase, not a quote, and show me the difference.
- If a claim has no source, say so plainly and mark it unverified. Do not fill the gap with something that sounds right.
- For every number and every date, ask what moment it was true and whether it is still true now.
Return two lists: claims you verified and how you verified them, and claims you could not verify. Change nothing in my draft. Recommend nothing I did not ask for. Tell me only what will not hold.
The first time you run it, watch for the sentence you were sure of, the one with no source behind it. The prompt cannot verify that one for you. It marks it unverified and hands it back, and closing that gap is the capability itself. The prompt does not replace your discernment. It gives your discernment its reps. You are learning to distrust the sound of your own work, which is the beginning of being able to check it.
What have you already published under your name without checking, because you did not yet know how?
This is Part 6 of Thinking With AI, a multipart series where I chart my journey from AI Consumer to AI Architect. Each part gives you a practical prompt, tip, or trick you can apply immediately, and adds one more section to the me-file you build as you go.
Part 1, My AI Called Me Gary - the 5 output checks and the first step in building your me-file
Part 2, The Explorer Stopped Getting Paid - the pitfall of information hoarding, and what happens when AI assumes a core part of your identity
Part 3, I Gave My AI a Map of My Weaknesses - naming your patterns out loud, and the Patterns to Catch section
Part 4, My Best Prompt Is 12 Words Long - the interview prompt that closes scope fast
Part 5, Your AI Is Guessing How You Think - why a fresh session guesses how you think, and the How I Think section
If you are new to the series, start with Part 1, where my AI called me Gary and taught me the fifth check I run on every output. That one also starts your me-file.