Essay · July 2026

The First Thing AI Replaces Is the Question

AI replaces the question before it replaces the work.

The first thing AI takes from you is your question, long before it touches your job. Most of the worry goes to the answer, whether it is accurate, whether it is hallucinated, whether you can trust it. That is the second problem. The first one is quieter, and it is already here: AI replaces the question before it replaces the work.

You used to sit with a problem long enough to frame it, to work out what you were actually trying to find and what a good answer would even look like. That framing was the work; it was where the thinking happened. Now you skip it. You type a rough version of the question, the answer comes back in seconds looking clean and confident, and you are reacting to it before you ever decided whether it was the right question to ask. The asking disappears in the time it takes to type a prompt.

Across studies of how people actually use these tools, only about one user in ten treats the system as a sparring partner, arguing with it, demanding the counter-argument, pushing on the answer before trusting it. That group consistently outperforms both people working alone and AI working alone, and the edge holds even with smaller, cheaper models. Vivienne Ming, who has done as much as anyone to make the finding legible, calls them a different kind of user altogether. The other nine in ten take the first answer and move on.

Intelligence is not what separates them. What does is perspective-taking, curiosity, and the humility to genuinely wonder what you are missing, and every one of those can be taught. None of them is automatic, though, and the tool will not supply them. Left to itself, AI rewards the opposite, because it is built to be fluent and agreeable. It will hand you a confident answer to a sloppy question and never let on that the question was sloppy.

I have a method I have leaned on for years with people, and it works just as well on a machine. Four questions: "Help me understand," "Why not?", "What if?", and "What would it take to?" They pull out what a confident first answer hides. I hit this directly last year, using an AI system to sort out a course load for one of my sons. I had uploaded his school's curriculum guide, and the model still answered for a generic high school, inventing courses that did not exist at his. We went around three times before I narrowed the question to one line: based only on this document, which electives satisfy the requirement he still has to fill. The data had been there the whole time. The model needed the question to use it.

That is the pattern everywhere. The machine will produce a sophisticated answer to the wrong question and deliver it with all the confidence of the right one. The moment that taught my team the most was a polished analysis built on a question no one had stopped to examine. We had asked the wrong thing, and once we saw that, the fix was obvious. After that, we built the question itself into how we check the work, the same way we check the output.

So before you prompt, write the question out by hand. One sentence. Read it back and ask whether it is really the question you want answered. Then prompt. When the answer comes, push back at least once, and if it cannot survive a single round of pushback, it was not ready, and neither were you.

The machine will answer anything you ask. It will never tell you that you asked the wrong thing. That part of the work never left you, and almost everyone has quietly set it down. The few who keep asking are the ones this era is rewarding.