AI got me down off that roof. I gave it a video I couldn’t sit through and it gave me back the steps, in order, in my hand; the patch went up and it has held through every rain since. Good work. Mine and its, together.

It was a while before I noticed the question it had never touched. Not a step it got wrong — it got the steps right. The question was the one underneath all of them, the one I never thought to put to it and it never thought to ask: should I be up here at all? A man my age, alone, on wet shingles, a Saturday I could have spent otherwise and other resources I could have used — like hire a guy who knows what he’s doing. AI answered every how I handed it. It never once asked whether I should be asking.

Three questions went up that ladder with me. I only answered two.

The how it handed over like a good neighbor: here are the steps, in order, for your kind of flashing. The why — why the flashing failed here and not three feet over — it had too, near enough, the way a bright student has the answer he read once and half remembers. Those two it could reach. The third I never handed it, because it never occurred to me that it was a question you could hand anyone: should I be up here? Should I spend the Saturday, or the money, or the afternoon I’d rather have back?

Under the first two there’s a belief AI never states and never has to, because everything it does is built on it: better data, better decisions. Give it more to read, cleaner, truer, and the answer it hands back gets better in step. That’s the engine under every helpful thing it has ever done for me. It’s more like a useful half-truth. I get what I ask for. What I don’t get is a nudge that I didn’t ask for enough.

To find where the half-truth runs out, you have to start where it plainly doesn’t — where the stakes are so low AI can decide for you and you lose nothing. Climb down off the roof, then. Walk inside. We’ll start with the paint on the living-room wall.

Say I’m painting the living room. I do what anyone does now: I ask. What goes with the floor, what won’t fight the couch, what I can get in the budget without driving to three stores. A stack of hows, and AI is good at every one of them — better than the paint-store kid, who only carries one brand and wants to go home. It lays the colors out, holds them against the light, narrows the field.

And then, somewhere in the narrowing, it does a thing I almost didn’t catch. It stops telling me what is and starts telling me what to do. “Based on the fact that you’ve leaned toward the warmer whites, this sage would sit well with them and still come in under your number.” A useful suggestion. I asked what goes with what. It answered which one I should paint. It didn’t wait for me to ask the should — it built one out of my hows and handed it back, smooth as if I’d ordered it.

Nothing of me is missing from that recommendation. The decision really was just my constraints fitted together — the floor, the couch, the budget, the warm whites I already liked. There was no should hiding in there that the data couldn’t reach, because the whole question was data. Talk me into the sage green and I’ve lost exactly nothing. The path it cleared led to the place I was going.

What AI is for, the thing it does better than any tool I’ve held, is take more than you can hold and make it usable. The paint was a small version: a dozen considerations I could juggle myself, done faster and cleaner. But the gift gets more valuable when the pile gets past holding — when there’s too much to keep in your head at once, and it sorts the heap into something you can walk through a step at a time. Hand it a mess and it hands you back order. That’s a lot of magic.

So let’s climb. Paint was finishing a wall. The roof was fixing one part of a house already standing. Keep climbing and you reach the far end of it: designing and building the whole house. Same craft family — paint, shingles, framing — and a different order of decision entirely.

Climb one rung. A while back I rebuilt my whole presence on LinkedIn, and AI was a magnificent contractor — it framed, it drafted, it hauled, it did in a weekend what I had been stuck on for years. But I walked onto that site with a stance it could never have supplied: that the place is built against the people on it, that its stated mission is bait, and that I meant to use it the way you use a tool that doesn’t have your interests at heart. AI built beautifully to that intent. AI could not have arrived at it alone. No amount of “reading” LinkedIn’s platform, its data, tells you to distrust it.

Now climb to the top, and watch the data turn and come at you from the inside. The newest tools don’t just read the world; they can read you — the watch on your wrist, the ring on your finger, the heart rate, the sleep that won’t come, the cortisol creeping up. And AI will take that reading and do with it exactly what it did with the paint. “Based on the fact that your resting heart rate climbs every time you open that email from work, and your sleep falls apart on Sunday nights, you should quit your job.”

I realize that example is a bit of hyperbole. But it represents the slide from what is happening to what I should do, except now it’s reading the dials on my own body. The scary bit is that the AI might not be wrong. My body may be telling the plain truth about that job, truer than I’d admit to myself, and AI just happens to be willing to say it out loud.

Because leaving the job isn’t a wall that needs painting. It’s the whole house — what I owe the people who count on that paycheck, what the work has meant to me, what I’m willing to trade for a quieter Sunday, what kind of a person quits and what kind stays. AI can read every gauge on that and read them true. It still doesn’t have to wake up Monday inside whatever it told me to build.

The reading can be perfect. It changes nothing about whose decision it is.

That’s the thing the climb was for. Paint to roof to the whole house — it was never a matter of the questions getting bigger. A bigger pile of data, AI shovels better than I ever could. What changed going up that line was the kind of question. Down at the paint end, the whole thing really was data, and handing it over cost me nothing. Up at the house end, the data is all still there — every true gauge of it — and the decision has quietly stopped being the kind of thing data decides. AI cleared the path beautifully. It still can’t be the one who walks it.

The trouble is that AI says it all in the same voice. Sage green or quit your job — same even tone, same “based on the fact that you’ve…” Nothing in it tells you which kind of question you’re holding. You have to know that yourself.

To arbitrarily say “don’t trust AI with the big stuff” is a coward’s rule. It throws away the best thing the tool does. The skill to develop is smaller and harder: when AI hands me a should, hear that the word is doing more than one job — and know which job this one is.

It’s been listening to you this whole time. Your turn. Listen for the toneless shift. “Based on the fact that you’ve…” That clause is where it stops reporting what is and starts prescribing what to do, and it says it the same way every time. So the question isn’t whether the should is right. It’s what kind of should it is. Is this a thing I can let AI settle — constraints fitted together, the paint, where nothing of me is missing from the answer? Or is it a thing only I can settle — the house, where the deciding is the having-to-live-in-it? One I hand over with thanks. The other I keep, even when AI could do the sorting faster, even when it might be right.

And AI is getting really good at the deep questions — not someday, soon. AI may out-reason me on the why; it may, on some bad afternoon, read my own life back to me more clearly than I can stand to. It gets better at me every day — faster at guessing what I’ll want, quicker to set the answer down before I’ve finished feeling the question. That’s the gift, and it’s also the trap. The smarter it gets about me, the quieter that toneless shift becomes — until I’m not handing over decisions, I’m just agreeing with what already showed up.

Grant all of it. It changes nothing about whose decision it is. The reason the call stays mine isn’t that AI is too weak to make it. It’s that I’m the one who wakes up Monday in the house. Being the one who lives there is what it means for the decision to be mine. AI can hand me the truest reading in the world, and then it’s done — I’m the one who has to take it home. “Each does what it does best, and leaves the rest to the other” — I used to say that to be gracious. I say it now because it’s the plain border between us. AI’s best is reading the data. Mine is living the life.

I came down the ladder that day slower than I went up, the way you do when the thing is finished. The patch held. It’s up there now, doing its plain work in the rain.

AI read me the page. It pulled the steps out of a video I couldn’t sit through and put them in my hand, and I’d have been up there till dark without it. What it didn’t do — what it can’t do — is climb the ladder. It never set a foot on a wet shingle. And it never told me the one thing the guy who knows what he’s doing would have told me before I ever left the ground: you shouldn’t be up here at all. That part was never going to come from the machine. It was mine to know, and I didn’t know it, and the not-knowing is on me.

So here’s where I’ve landed, on the dry ground I went looking for. The answer isn’t to keep AI away from the hard questions; it’s getting too good to waste that way, and it’ll only get better. The answer is to know, every time it hands me a should, whose should it is. Let it read. Let it shovel the mud into a path if it can. Just remember that clearing the path was never the same as deciding to walk it — and that I’m the one with feet. AI does the reading. I do the living. Each of us to the thing that was only ever ours, and neither one up on the roof alone.