It’s officially confirmed. The suspicion many of us held is now backed by data: using AI to code can actually make you a worse programmer. Anthropic recently concluded a landmark trial pitting two groups of developers against each other—one with AI assistance and one without. The results were stark. The developers using AI scored a staggering 17% lower across the board.
This isn’t necessarily a catastrophe. Since the dawn of time, humanity has embraced technology to enhance its capabilities, often at the expense of a core, manual skill. Think of hunting, sewing, farming, or countless other abilities that were once essential for survival. We’ve traded raw skill for technological leverage.
The reason for this article is to address the belief that programming is somehow different, that it exists in a special category immune to this trade-off. Well, it’s not. The study definitively shows that our coding skills are vulnerable when we rely too heavily on AI assistants. And while Anthropic’s study focused on programming, it doesn’t take a huge leap to see this trend applying to other skills like writing, critical thinking, and problem-solving.
Dissecting the Anthropic Study
So, what exactly did Anthropic do? They examined how quickly software developers could learn a new Python library, comparing a group with AI assistance to a control group without it. The goal was to determine if using AI affected their comprehension of the code they had just written.
The findings were clear: AI assistance leads to a statistically significant decrease in what the researchers termed “mastery.”
The study’s design was straightforward:
- Warm-up: A 10-minute coding task where neither group used AI.
- Main Task: The treatment group was allowed to use an AI assistant, while the control group was not.
- Post-Task Quiz & Survey: A quiz to test their understanding of the concepts they just used. Neither group had access to AI for this part.
The only variable in this randomized controlled trial was the use of AI during the primary learning task. The results, while perhaps not shocking, are incredibly telling. Using a tool that improves leverage and eases the workload naturally diminishes the core skill. When was the last time you used a bow and arrow to catch dinner? We don’t have to anymore, and consequently, most of us are not very good at it.
The Illusion of Speed vs. The Reality of Skill
The study’s data revealed a fascinating contrast.
- Time to Completion: The AI-assisted group finished the task in about 23 minutes. The non-AI group took approximately 1.75 minutes longer. Statistically speaking, this small difference was not significant enough to declare a major speed advantage for the AI group.
- Quiz Scores: This is where the real story lies. The group that used AI scored an average of 50% on the mastery quiz. In contrast, the group that coded manually averaged between 65% and 66%. That 17-point gap is the equivalent of two full letter grades.
It’s important to note that Anthropic isn’t dismissing AI entirely. They expressly pointed out that their own previous research has shown AI can help people complete some work tasks up to 80% faster. The crucial difference? The previous study measured productivity on tasks where developers already had the relevant skills. This new study examines what happens when people are learning something new.
The takeaway is that using AI during the completion of a task appears to hinder the ability to learn while doing it. This is as groundbreaking as a study showing that getting stabbed is bad for your health. Of course it is. Learning requires struggle. You have to approach a problem from different angles, try things that don’t work, and engage in the hard work yourself. How else will your brain form new connections?
Imagine you’re at the gym doing a bicep curl, and a friend starts helping you lift the weight. Would you expect your muscle to grow as much as it would if you did all the work yourself? Obviously not.
A Tale of Two Growths: Horizontal vs. Vertical
Despite the obviousness of the results, there are solid takeaways. AI improves your leverage, but it does so by eroding your core skills. This introduces two forms of production: horizontal and vertical.
Horizontal Production is when you take an established, working approach and simply multiply it. Imagine a business that creates a popular new soap. Horizontal growth is buying more factories to produce more of that same soap. You aren’t innovating; you’re scaling. In programming, AI has supercharged our ability to grow horizontally. We can produce more well-documented, well-established products, but this comes at the cost of vertical growth.
Vertical Growth is our ability to apply human ingenuity to a problem to develop a fundamentally better way of doing things. It’s not building more soap factories; it’s inventing a new process that allows one factory to do the work of a thousand. This type of growth is incredibly difficult and occurs in unpredictable bursts.
AI is a tool for horizontal expansion. It makes the most sense to use it when you already have a solid foundation of skill. With a strong base, you can produce more articles, more code, more applications, and more websites at a faster rate. But this focus on horizontal expansion means you aren’t engaging in vertical expansion—you aren’t working on the core skill itself.
The Strategic Trade-Off: What to Outsource to AI
Rather than lamenting this reality, it’s better to make the trade-off knowingly. Since there is an erosion of core skill, the wise approach is not to rely on AI for everything. You must pick and choose.
-
Coding: As an economically valuable task, it’s a candidate for AI assistance, especially for repetitive or boilerplate work. If you can produce something marginally faster—even by just two minutes out of 25—it might make business sense. But be aware of the cost.
-
Personal Relationships & Critical Thinking: These are areas you should never outsource. How to communicate with a partner, how to navigate interpersonal dynamics, and how to make major life decisions are skills that define you as a person. No matter how good AI gets at producing widgets, these human elements aren’t going anywhere. Don’t pass off a conversation with your spouse to a chatbot.
Like it or not, AI will soon be better than us at nearly all economically valuable skills. If you build your entire identity around your ability to produce those skills and then let AI do it all for you, who are you?
You are the sum of your relationships. You are your character. You are what’s between your ears. If you’re going to outsource anything, outsource the economic part. Don’t outsource the you part.
The Evolving Definition of “Work”
The nature of work is changing. In a recent conversation, a friend raised a point that was both annoying and profound. I was arguing that with AI and robotics becoming superior in every way—thinking faster, working tirelessly—there might be no future careers left for people.
He countered, “I bet if you took a hunter-gatherer from 10,000 years ago, or even a farmer from 3,000 years ago, and showed him what you do all day, there’s no way he’d call it ‘economically valuable work.’ He’d ask, ‘Where’s the hunting? Where’s the gathering?’ To him, you’re just sitting at a computer twiddling your thumbs.”
His take was that we are already doing jobs that our ancestors would find abstract and meaningless. What’s the real difference between the jobs we do today and the ones we’ll be doing in 50 or 100 years?
It was an irritatingly good point. Who’s to say he’s not right? Hopefully, this article has at least made you think.