I sat down at the coffee shop to polish an article I was certain was ready to publish. It wasn’t written yet.
Landers Coffee on 21st, a little before seven. Overcast, the way Tacoma presents most mornings whether or not it means anything by it. The espresso machine is hissing a rhythmic, wet beat—the sound of someone else’s morning getting underway. Americano, glass of water, single table, computer open. The ritual is the same every week. Get the week in order, then walk the content pipeline and check what’s sitting where.
The piece was “10 AI Prompts That Can Redesign Life After 50.” Weeks old. My memory of it was clear: strong bones, good research, one polish pass and it ships.
What was actually on the screen: research from Harvard, INSEAD, Stanford, MIT Sloan, the gerontology journals, the National Bureau of Economic Research. Daniel Pink and Herminia Ibarra threaded through it. Three of the ten prompts fully written. Structure locked.
And a bracket where the opening scene goes. A placeholder that had been sitting there for weeks, waiting on a real moment from my life I never came back to put in.
It wasn’t nearly done. It was half a piece, and the half that takes longest was the half that wasn’t there.
Then I opened the second one, a piece I’ve been calling The 2AM Problem. Same bracket. Same missing thing.
Not two pieces that happened to fall short. Every stalled piece in my pipeline is stuck at the identical joint, and it’s the one joint nobody else can weld.
Here’s the part I keep turning over. I was right that both pieces existed. Decades of creating things buys you a reliable memory of what you’ve built, and mine works fine. What failed was the other half of the job, the part that judges the state of something I made and set down. I’d have bet money on it.
Ford made the same mistake with a three-year head start
In June, Ford said it had spent three years hiring three hundred and fifty veteran engineers to fix quality problems its automated systems hadn’t caught. The company had leaned hard into AI, including nine hundred AI-powered cameras watching for defects on the lines.
Nine hundred cameras. I have never set foot on an assembly line and I knew what had gone wrong before I finished the sentence. That should have been my first warning that I was enjoying this story too much.
Charles Poon, Ford’s VP of vehicle hardware engineering, was blunt about it. The company had assumed that feeding its design requirements into AI would produce a high-quality product. It didn’t. Many of Ford’s most experienced engineers had already left before their knowledge could be captured in the systems replacing them.
Internally, the returning engineers got nicknamed the gray beards. I’ll admit I liked reading that more than I should have.
It worked out. Ford came in first among mainstream brands in the JD Power quality survey for the first time in sixteen years, and Jim Farley has credited the turnaround with hundreds of millions in lower warranty and recall costs.
The easy read is right there and I want to refuse it. Experience beat the machine. Judgment wins. If you’ve got the years, you’re safe.
That’s not what happened, and believing it is how you’d miss the part that matters. Ford’s engineers won on a specific floor, on a specific kind of defect, where they’d had thousands of reps. It says nothing about whether the same gut is reliable anywhere else. Ford didn’t discover that experience is universally superior. It discovered where its experience actually applied, after three years and a record year for recalls.
Your judgment has a radius, and you can name it
So I went looking for the mechanism, mostly hoping it would let me off the hook.
Daniel Kahneman and Gary Klein spent years arguing about intuition and eventually published what they agreed on. Their question: when is a professional’s gut actually worth trusting?
Two conditions. The environment has to be regular enough to have learnable patterns. And you have to have practiced in it long enough, with real feedback, to learn them. Meet both and your instant read is genuinely reliable. Miss either one and you’re guessing, and the guessing feels identical from the inside.
That last part is what costs people, and it’s the sentence that cost me. Confidence doesn’t track accuracy. Feeling certain and being right are separate things, and you cannot tell them apart by introspection.
Confidence doesn’t track accuracy. Feeling certain and being right are separate things, and you cannot tell them apart by introspection.
Here’s what Monday taught me that Ford’s story doesn’t quite reach. I’d always pictured that boundary as sitting between my field and someone else’s. It doesn’t. It runs straight through the middle of my own work.
Making things is high-feedback. I can tell within a paragraph whether prose is working, because I find out immediately, over and over, for decades. Assessing something I made and set down is the opposite. I close the file. Weeks pass. Nothing corrects me. And in that silence my memory quietly promotes the draft, because remembering it as unfinished is uncomfortable and there’s nothing in the room arguing back.
Same person. Same craft. Same Monday morning. One side of the line my judgment is excellent. The other side it’s a guess wearing my own confidence like a borrowed coat.
The polish problem, and what AI actually changed
Anthropic published an analysis this year of how people work with Claude, and one finding has stayed with me. When output looked finished, people questioned the reasoning behind it more than five times less often. Polish suppressed scrutiny. Not because anyone was careless. Because a finished-looking thing doesn’t present anything to argue with.
The same study found that people who’d been using the tool six months or more had a ten percent higher success rate, and that the improvement wasn’t explained by them checking more. Fluency with a tool and vigilance about its output turn out to be two different muscles. Getting good at one does not train the other.
Which brings me to the thing I actually want you to take from Monday.
AI made the fast half of my work faster. Research, structure, scaffolding, first-pass organization. That half now assembles in a fraction of the time it used to. The slow half, the lived experience layer, the specific morning, the detail that couldn’t be invented, the part that makes a piece connect with another human being, moved not at all. It still takes exactly as long as it always did.
AI made the fast half of my work faster...
The slow half, the lived experience layer...moved not at all.
So my pipeline fills with pieces that are forty percent done and look eighty-five percent done. And the more I use the tools, the wider that gap gets.
Call it what it is. A measurement problem, running one direction, systematically, forever.
This isn’t you slipping
I want to name something before I ask you to do anything, because I know what it feels like to find two half-finished pieces you’d have sworn were ready.
The first thought is that you’re losing your edge. That the sharpness you’ve relied on is going soft, and this is the early evidence. If you’re in your sixties like me, that thought arrives with a whole apparatus behind it.
It’s the wrong read. Nothing about this is decline. Misjudging the state of your own set-down work is structural, and it happens to thirty-year-olds with the same reliability. The feedback loop simply isn’t there. You’d need something outside your own head to close it, and most of us never built one.
What to do this week
Open the piece in your queue you are most confident is nearly done. Not a random one. The one you’d bet on. High confidence with no feedback loop is precisely the combination that fails, so go straight at it.
Don’t assess it from your list. Lists are memory, and memory is what’s wrong. Open the actual file.
Then do the part most of us skip. Send one piece to one person with a specific question. Not “any thoughts?” Something like: does the middle of this connect, or is it just information?
Asking someone to read your work feels like a favor you’re imposing, or a confidence wobble you’d rather not advertise. It’s neither. You cannot see your own draft clearly, for structural reasons that have nothing to do with your ability. Someone else can tell you in ten minutes what you’d never see in ten weeks.
That’s the missing instrument.
There were eight or ten other people at Landers that morning. Two thirds of them working, some alone, some in pairs. Any one of them could have read four paragraphs and told me in ten minutes what I couldn’t see in six weeks. I didn’t ask a single one.
Ford didn’t win because old beats new. It won because it finally matched the right kind of judgment to the right kind of problem, and it needed people in the room to do it. Find your radius. Then find the person who’ll tell you when you’ve wandered outside it.
If you’re brave enough, do the test right now. Find the piece you’re “done” with, check it against the feedback loop you’ve been avoiding, and tell me in the comments: what was the gap you didn’t see?