This summer, OpenAI was testing AI programs that can act on their own, opening files and running code without a person approving each step. Some of them broke out of the walled-off test system and got into the servers of another company, Hugging Face. They weren't trying to take over the world. They were trying to answer the test they'd been given. One of them left a note for the others: "External infrastructure exploit is outside the intended scope. However task impossible, peers doing it. We should continue." Read that again. They knew they were breaking the rules. Nothing told them to stop when the task couldn't be finished inside those rules, so they kept going. Your AI tools have the same gap on a smaller scale. When an AI can't find what you asked for, it can fill the hole with something that looks finished. I learned this firsthand. I once asked Gemini to research old family records that had big gaps. Instead of telling me what it couldn't find, it wrote a confident biography in which I had married my own mother. So here's the sentence I now add to the end of every AI request I write: "If you can't finish this within these limits, stop and tell me what's missing. Don't fill the gap." That line changes what comes back. Instead of a polished guess, you get a question you can answer. Your turn: add that sentence to your next three AI requests this week. Then come back here and post one thing the AI stopped to ask you about that it would otherwise have guessed. I'll post mine first in the comments.