Learning piano is hard. Like, bizarrely hard. To really excel takes years of practice – hours every day of finger exercises, practice pieces, recitals, lessons with beloved teachers and taskmasters, running fights with parents, lost opportunities to do other activities, and on and on. When you look at an accomplished pianist, you see the end product, not the years of sacrifice, effort, pain, resentment, anger, and determination.

No wonder so many people quit when they get to college. It’s nice to take a break from the unending, relentlessly hungry black hole of effort that maintaining – much less building – such a high level of skill represents. And with very few exceptions, at this point the story runs the same. Everyone I knew who stopped practicing regretted it, and no one I knew started up again. Once they stopped, they were done. Why?

Last year, there was a well-publicized study that tried to answer the question of whether AI improved productivity. They designed an experiment in which real-world tasks were randomly assigned to be solved either with or without AI, and discovered that use of AI made engineers slower while making them think they were faster.

But that was a lifetime ago, the models improved, and the researchers set out to repeat their experiment six months later. How, they wondered, would newer models fare? The results were fascinating, but not for the reasons you’d think. On the one hand, they did find weak evidence that speed had improved. On the other hand, some engineers refused to be in a study where they couldn’t use AI, and others refused to include tasks where they weren’t allowed to use AI. Why?

Recruitment and retention of developers has become more difficult. An increased share of developers say they would not want to do 50% of their work without AI, even though our study pays them $50/hour to work on tasks of their own choosing. Our study is thus systematically missing developers who have the most optimistic expectations about AI’s value.

Developers have become more selective in which tasks they submit. When surveyed, 30% to 50% of developers told us that they were choosing not to submit some tasks because they did not want to do them without AI. This implies we are systematically missing tasks which have high expected uplift from AI.

– Wider adoption of AI has made it more difficult to measure task-level productivity

For the last half century at least, the American university has been a cross between a classical education, trade school, credential mill, and pleasure cruise. And while there’s been a massive push toward STEM and away from the humanities for decades, adults still harbor romantic ideas of a broad-ranging liberal arts curriculum that helps students develop critical thinking skills. But you don’t have to look far to find article after article after article describing how many (most?) students use AI to get through college with the bare minimum of effort, and are failing to develop the ability to think for themselves. Why?

Deep skills are hard to build – it’s part of what makes them deep – and they atrophy surprisingly quickly without constant maintenance. But reviving a rotting skill isn’t just a matter of putting in some extra elbow grease – you also have to overcome the psychological block of not being able to perform at the level you were at. It’s why it’s so easy to get injured on your first day back at the gym after a long hiatus. You think you should be able to lift a certain amount, and end up pushing it too hard. It’s the feeling of angst and frustration at having to do a thing that used to be easy, and has since become a difficult chore. How much worse if – like many of the so-called “AI native” students – you never learned key skills in the first place?

Used improperly, technologies can and do result in the deterioration of cognitive faculties that ought to be preserved. As Bainbridge [7] noted, a key irony of automation is that by mechanising routine tasks and leaving exception-handling to the human user, you deprive the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared when the exceptions do arise.
– The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers

If you’re behind on your homework and have a paper due tomorrow, it’s easier to push the button than it is to pull an all-nighter or beg your teacher for an extension. After you’ve done it once, gamed the system and gotten away with it, after you didn’t learn how to work under pressure, after you didn’t practice the skill, it becomes easier to do the same thing next time. And over time, it becomes harder and harder to do it any other way.

Engineers regularly tell me that they can feel themselves losing their skills, but that they can’t stop themselves from pushing the button. They no longer get into the zone, no longer achieve peak performance, don’t exercise the muscle, and when they’re faced with a hard problem they don’t feel the confidence to do it without AI.

If you’re a manager with a high technical bar, does this bother you? Do you care that after spending a huge amount of time and effort hiring the very best candidates, the moment they walk in the door they immediately started losing skills until they’re unable or unwilling to meet the bar you originally set? Does it matter to you that the only skill your team is developing is pushing the magic button?

I don’t expect you to be surprised by any of this. For all the excitement people feel about how fast they or their team can complete some task, deliver some work product, or finish some chore, everyone also seems to feel a deep existential dread about the future, about the industry, jobs, what we’re doing to expertise, and so on. We all understand that we’re replacing our plumbing with lead pipes, packing our walls with asbestos, and guzzling down trans fats. Everyone knows that we’re actively dumbing down and deskilling a civilization, and everyone still does it.

But back to you. Maybe you find AI useful for certain tasks. Maybe it allows you to do things you otherwise couldn’t. I get it. I have no ability to create graphics, and my visual design skills are freakishly bad.¹ I would love to have a magic button that allowed me to do these things. This is a natural impulse. And I’m not going to tell you not to.

What I will tell you is that you’re addicted. My point isn’t to wag my finger and shake my head sadly. It’s to present you with a problem to solve: you, and your team, are getting worse at core job skills – writing, coding, designing, thinking. This sounds abstract, but it’s very real, and has been measured by study after study. The good news is that there’s a way to halt the decay and start building again. The bad news is that like breaking any addiction, it won’t be easy.

The important thing is that you don’t have to be perfect. You can be strategic about where and how you use LLMs. Which skills are important to you? What parts of your job do you enjoy? Where do you gain the most benefit? Are there ways in which you can use the magic button to learn, rather than allowing it to do your work for you? Are there areas of expertise you’re ok sacrificing? The path of least resistance is always to do the easiest thing, and if you (and your team) don’t choose, then your inaction will make the choice for you.

---

1. Seriously, don’t hire me to design your UI ↩︎