How often does AI get it wrong? Mine, nearly 1 in 5 times. A second AI checked 9,616 pieces of my AI’s work in five weeks and sent back 1,752 before I used any of them.
The turn.
AI does most of my research now. It reads other businesses’ ads and pages and writes up what they sell and who buys. I can’t check all of that myself.
So every piece goes to a second AI first. It holds the work up against its source and sends back anything that doesn’t match. Only then do I use it.
The proof.
These numbers come from my agents’ work log, 5 September to 12 October 2026.
The checks looked at 9,616 pieces of work. 7,864 passed the first time and 1,752 were sent back: 18%, nearly 1 in 5. A second AI sent back most of the returns; automatic checks caught the rest, like a number that wasn’t in any source. Then I read 60 of the mistakes one by one, picked at random, to see what they were.
It isn’t only mine. When 22 public broadcasters checked over 3,000 news answers from ChatGPT, Copilot, Gemini and Perplexity, 45% had at least one significant issue (EBU and BBC, October 2025).
Lesson 1. I ask what it left out.
About half the mistakes I read were gaps: 31 of 60. A buyer group, a product, or a detail that was in the source but not in the write-up.
A missing fact is harder to spot than a wrong one, because nothing on the page looks off. So I ask for it directly.
Try this, if it fits. When AI sums up your reviews, could you ask it: “What did you leave out?”
Lesson 2. I ask for the line behind each claim.
About 1 in 4 went past the source: 16 of 60. A Reddit comment said “We also use copilot for PR reviews”. My AI wrote “Their team uses a paid review tool.” Nothing in the comment said anyone paid.
If AI can’t show the line, the claim goes.
Try this, if it fits. Could you ask AI to quote the exact line behind each claim it makes?
Lesson 3. I send it back with the reason.
Over half were fixed on the very next try: 986 of 1,752. Each return says what was wrong, where, and what the source actually says. A vague “try again” gets a new answer with new mistakes. A specific note gets the fix.
Try this, if it fits. Could you tell AI exactly what was wrong, and ask for only that fix?
What I do now.
Every piece of research my agents write goes through this check before I see it, and the same habit runs in the AI automation work I set up for other businesses.
Want the steps in order? They’re in my free guide, is ChatGPT accurate?
