When a chatbot makes something up, we blame the chatbot. Fair enough — it made it up. But I analyzed 522,607 questions that people typed into ChatGPT, Claude, Grok, Gemini and Copilot, and the bot is not the only one to blame. It’s also you.
I found eight ways we slip up, and they are all the same slip: part of your answer is already in your question, and you don’t notice — but the machine does. It finds the answer sitting in your question and hands it straight back.
It is the photocopier you keep swearing at because there is a smudge on every single page it produces. The smudge is on your original. The machine is doing the one thing it is built to do, which is give you a faithful copy of exactly what you fed it — cleaner, sharper, five hundred words longer, and just as wrong.
Half a million questions, nineteen languages, 46,287 exchanges rebuilt into question-and-reply, and finally 9,000 judged one at a time. I wasn’t looking for the questions that ask, but for the ones that steer the answer. Why does the report ignore the obvious cost problem — that one steers. Whether there is a cost problem was never checked; it rode in as furniture, and the machine built on it.
When a question carries a claim inside like that, the chatbots let it stand in 19 out of 20 replies.
This is the third piece in a series about a large pile of other people’s chatbot conversations. Part one was about what people gave away in them without meaning to — pension numbers, visa dates, a man asking whether this could stay between us, on a page anybody can still open. Part two was about the rabbit holes the machines pulled people into, including thirteen hundred messages of somebody’s imaginary physics.
How do we slip up?
Eight mistakes, each with its own fix below. But the fixes are all one habit applied in eight places, and the habit comes down to five rules.
