I Appreciate AI. I’m Also Grateful My Career Came Before It Arrived.

How much of what I’m good at came from times when I was stuck, wrong, surprised, or embarrassed?

I’ve been asking myself that a lot lately.

My best guess is: most of it. But I can’t run the control group on my own life. I’ll never know who I’d be if the answers had always been one click away.

A short list of my mistakes

What I can say is that much of my professional judgment came from getting things wrong.

Early in my career, I drew conclusions from preliminary data I was thrilled about and showed my boss. Before I could go back and check my work, he had handed it to marketing, and marketing had called The New York Times. Uh-oh.

Another time, I believed my claims analysis had discovered a new, widespread injury among manufacturing workers. It turned out to be a quirky procedure code used by the team’s favorite chiropractor. (I still remember it: "other ill-defined dislocations of the cervical vertebrae.")

Fresh out of academia, full of stat jargon, I once presented complicated (and immensely important) findings to several business leaders and got blank stares. I lost credibility, not because the analysis was wrong, but because of how I presented it.

And to put myself through grad school, I worked as a statistical advisor. That gave me hundreds of examples of what can go wrong. I watched when others dumped hundreds of variables into a model and let the math decide what mattered. One study of why college students drop out suggested that the most important factor was school colors. (Um, purple and yellow are probably not the reason).

Each of those left a reflex behind.

Double check before you share. Ask about the odd code. Translate before you present. Distrust a result that makes no sense.

Years later, when junior analysts told me an analysis showed “no relationship,” that reflex made me ask them to look again in a new way. Sure enough, the relationship was there, hidden in another factor, which made the story was even more meaningful.

None of that came from watching someone else get it right.

Learning researchers have a name for this: “desirable difficulty,” a term coined by psychologist Robert Bjork. It describes struggle that builds lasting skill instead of just slowing you down. Watching someone else do the work teaches you what good looks like. Doing it teaches you what to do when something looks good and isn’t.

What the evidence says so far

So does AI remove the struggle?

Research says it depends on how we use it.

In a randomized trial with nearly 1,000 math students at a high school in Turkey, students who practiced with a standard version of ChatGPT scored 48% higher on practice problems than students without it. When the AI was taken away for the exam, they scored 17% lower than students who never had it.

A second group used an AI tutor that gave hints instead of answers. They scored 127% higher in practice, and as well or better than students who never had AI.

Then there’s a two-year trial of Khan Academy’s AI tutor in 18 Tennessee middle schools. It was configured to coach rather than give answers. Nearly every student tried it, but few used it regularly.

In the researchers’ words: “Seeking help with one’s own confusion remained a choice, and most students declined it most of the time.”

AI can protect the struggle or remove it. So far, given the choice, most of those students skipped the coaching that required more effort.

I wonder, would a middle school version of me do the same?

I realize the rote work wasn’t empty

I appreciate AI. It pulls papers from across the web in minutes. It merges messy datasets into one. It runs analyses I can request without picky, esoteric syntax. It compares and summarizes huge amounts of desperate information. It has undoubtedly made me more productive.

Perhaps I notice its impact more than most, because my career began in the 1980s, before fax machines, personal computers, or the internet. My dissertation involved months in the library "stacks" (a nice word for several floors of dusty old journal collections below ground), dimes fed into the copy machine, data entered by hand, and punch cards carried to the computer center.

Easy to call that drudgery. But every dime forced a decision about what was worth keeping. Entering data by hand meant I saw every odd code. Each punch-card run took hours, so I thought hard about every line before submitting it.

One more challenge shaped my dissertation: I needed a time-series technique I wasn't sure existed. A professor hinted that economists might have one. Digging through their stacks, I found Granger causality. That hint worked like the tutor in Turkey. Had AI simply handed me a method, I'd never have learned a technique that has come in handy ever since.

Similarly, a clean, automated summary might have quietly reclassified that chiropractor’s code, and I’d have missed the lesson. The decisions still get made. Except they get made without you.

The questions I’m left with

Which struggles shaped you? Which of them would you hand to AI today without a second thought? And what lesson might you miss by doing so?

And if you lead people, which ones are they skipping right now, before they’ve had the chance to make their own mistakes?

I was surrounded by people who encouraged me to work through frustration and learn. Today, I am grateful I didn’t have an easy alternative.

AI can be used to help the learning process or to skip it.
Which one we choose is up to us.