Learning with AI

· 4 min read
#AI#learning

If you had to acquire a new programming skill today, you’d reach for AI. But here’s the thing — using it wrong might mean you never learn the skill at all.

A recent Anthropic study centered on this issue: how does the use of AI tools affect the ability to develop long-term skills when working with a new programming language, library or framework?

To conduct the study, they created a simple task: write two asynchronous functions using Trio, a python concurrency library, and divided the participants in two groups. The non-AI group was asked not to use AI assistance tools to complete the task, while the AI group was given an AI assistant that could provide the working answers to the task if requested.

What they found was a tradeoff: the faster AI got you to a working solution, the less you learned.

The most revealing part of the study was how participants interacted with AI — and how those patterns predicted learning outcomes. Within the AI group, there were 6 interaction personas, all of which I’ve been guilty of being at one point. The first 3 personas scored poorly, while the last three demonstrated high comprehension when quizzed:

  1. The Vibe Coder: participants in this group used the AI assistant to write code and complete the task. These were the fastest to finish and ran into little to no errors.
  2. The Reluctant Adapter: participants in this group started asking questions and tried to complete the task on their own, but after 1 or 2 questions, delegated all code writing to the AI assistant. Because they never fully committed to understanding, they scored poorly on the quiz.
  3. The Lazy Debugger: participants in this group relied on AI to debug or verify their code. Because this group asked the assistant to solve their problem, rather than to understand it, they had a lower comprehension score when tested. They were also slower at completing the two tasks.
  4. The Curious Vibe Coder: the only difference between participants in this group and the Vibe Coders, is that after the code was generated, they asked the AI assistant follow-up questions to improve understanding. They weren’t much faster than the non-AI group at completing the task, but demonstrated a high level of understanding on the quiz.
  5. Show Your Work: participants in this group composed queries where they asked for code along with explanations of the generated code. This group took longer to complete the task and was the second-highest scored among the groups.
  6. Old School: participants in this group only asked the AI assistant conceptual questions, and relied on their improved understanding to complete the task. They ran into many errors, but solved them independently. This was the fastest among the high-scoring patterns, and the second fastest to complete the tasks after the Vibe Coders.

Downstream Consequences

The study found real downstream consequences to low-comprehension AI use. Code that was generated but never understood still needs to be maintained. When the time comes to update, refactor, or debug it, the developer who wrote it won’t know where to start. They’ll be starting over, hitting the same walls, and reaching for AI again — compounding the problem.

This is especially true for junior developers, who are still building the foundational instincts that make debugging intuitive. If those instincts never form, the reliance on AI doesn’t shrink over time — it grows.

Should you stop vibe coding?

No. You shouldn’t. But it’s important to stay conscious of what you’re offloading - because what feels like productivity is borrowed understanding you’ll eventually have to pay back.

Out of the six personas, the one that will teach you the most is the Old School approach, with Show Your Work a strong second. But that can be a large jump if you’re new to programming. The Curious Vibe Coder is the best persona beginners can adapt to, and vibe coding tools make it even easier with options like Claude Code’s Learning and Explanatory mode or ChatGPT’s Study Mode. Asking AI assistance to explain the code they’re generating will help you pick up the necessary skills along the way to maintain your code in the future.

I’ve been all six

The results of this study align with my own personal experience using AI assistants. There have been a few times where I’ve built things off vibes, and when I come back to them, I have no idea what they are and how they work. On the flip side, the things I’ve built “Old School”, treating AI as a teacher and asking questions to deepen understanding, are the things I can come back to and not only understand, but even explain to others.

The bottom line

AI assistants are amazing tools, and they can get you from 0-100 in a short amount of time. But know that there’s always a price if you’re working with something new: the faster AI gets you to a working solution, the less you learn. If you want to be able to maintain your projects in the future, instead of contributing to the rise of abandon-ware, take the time to learn something along the way.

Sources