Good Enough Got Too Easy

At Apple, I over-engineered a lot. In April, I wrote about how AI made it worse. A review agent would always find one more thing. Fix it, run another review, find something else. I had given my inability to leave things alone a collaborator that never said no and always kept running.

I’m better about stopping now. Mostly. I can still take a task that’s in good shape after 30 or 45 minutes and spend the rest of the evening making it slightly more like the version in my head. The improvements get smaller. The token usage remains reasonable.

Then I read a new working paper where AI had the opposite effect.

The researchers recruited 219 people in China for two illustration tasks, each capped at 150 minutes. Everyone did the first without AI; 168 received AI access for the second, while 51 continued without it. Professors and industry pros rated snapshots of the work every 15 minutes.

Without AI, work sessions clustered at the time limit (no surprise, people tend to fill the allotted time). With AI, another peak appeared around 60 minutes, right where the quality curve flattened. Average quality was higher with AI at each checkpoint: in 15 minutes it reached the level that took an hour without AI. Yet of those offered AI, 77% spent less time, and 38% turned in lower-rated work than on their first task.

Add it up, and the average final score barely moved. People took the gain as time back.

The limits matter. This was one kind of creative task, with cash prizes. Most participants were art and design students, the average age was 23, and the experiment ran in 2024. It’s still a working paper.

I get why people stop. Once something looks finished, every additional improvement has to justify itself. With AI, that conversation arrives much sooner.

Stopping there can be a perfectly sensible decision. You got what you came for, and you’ve got other things to do. A lower score from an expert judge doesn’t tell us whether the time saved was worth it to the person doing the work.

There’s a trap in the flat curve, though. If another hour of iterating barely changes the result, I may have reached the limit of that approach. I still need to decide whether the work does its job. The tool’s diminishing returns can’t make that decision for me.

Do you know what you’re leaving unfinished?

I’ve been building Lemon Cart, an iPad game for my son. After a few days, he was playing it. I love that this is possible now. An idea I wanted to make for him was in his hands.

I could also see how easily I might have kept him waiting.

On some of the artwork, I called it after about an hour. I could have kept generating versions and adjusting details. There was always another image I might like slightly more. For what I was trying to give him, I had enough.

Gameplay held my attention longer. I wanted to make sure he couldn’t get stuck.

A game that strands a toddler is a parenting problem. And I’d have no one to blame but myself.

That gave me something concrete to judge the work against. Another hour might buy very little. It might buy the difference between getting to keep playing and needing Dad to fix something.

I also wanted it to feel magical. A game not crashing is the floor. A cloud should rain, and the people react. A lemon drops off the tree. A water bucket sloshes when you tap it. Kids love these moments. So do the grown-ups. That extra time was totally worth it.

I could stop on the artwork and keep going on the gameplay without lowering my standards for the whole thing. The extra effort needed a reason.

Of course, spotting that reason depends on what you know. An experienced engineer might notice a failure someone new to software wouldn’t know to check for. The screen can look equally finished to both of them.

Experience also gives you more ways to keep yourself busy. I have enough of it to invent a very convincing reason for another review.

On a team, this gets harder. One person sees a harmless rough edge. Another sees future support tickets. And a project can always absorb another week.

At Apple, we ran the rock tumbler all the time. That was Steve Jobs’s image for a team grinding ideas against each other until they come out polished. We kept putting one question through it: which checks had to run on every iPhone that shipped? It sounds straightforward. It isn’t. Every check you keep runs millions of times. Every check you drop is a bet on design validation and customer usage.

We’d argue for a few days, work out which customer use cases truly couldn’t fail, agree on the trade-offs, and move forward. Did we over-analyze sometimes? Sure. But we got better with every rep.

Have that argument before AI hands you an impressive version. Decide which failures hold the release and which rough edges you’ll live with. Then show the work early to the people who will own it. The teammate who supports the feature may catch the missing behavior. Someone close to the customer may tell you the detail you’ve been agonizing over barely matters.

Then someone needs the authority to say we’re done. Agreeing on what done looks like also gives people permission to finish. An agent will keep producing reasonable suggestions for as long as you ask. A conscientious person will treat each one as an obligation.

I still believe in painting the back of the fence. I like beautiful work, including details someone may only notice when they’re wrong. I also know how easily I can turn care into another evening spent avoiding a decision or other work.

The tool can offer another version indefinitely. I have to know what I owe the person receiving it.

When I’m tempted to spend another evening on the game, I picture who will be holding the iPad.

Who is your extra hour for?

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free