Learning to code in the age of AI: what you still need to know how to do yourself

An AI coding tool can write you a function in a few seconds. Does that mean you no longer need to learn how to code yourself? Three studies, published between 2025 and 2026, give precise answers, and not always the ones you’d expect.

A study that breaks an intuition: faster doesn’t mean faster

In July 2025, the research organization METR published the results of a controlled experiment with 16 experienced open-source developers (averaging 5 years of experience on the projects studied), who worked through 246 real tasks, sometimes with AI tools (mainly Cursor Pro with Claude 3.5 and 3.7 Sonnet), sometimes without. The result: allowing AI actually made completion 19% slower, not faster.

The most interesting part isn’t even that. Before starting, these experienced developers predicted the AI would save them 24% of their time. And after finishing, even though the AI had genuinely slowed them down, they still estimated it had saved them 20% of their time. In other words: even seasoned developers misjudge their own speed when AI is in the loop, in a way that doesn’t correct itself with experience. Correctly judging whether an AI is actually helping you already requires a solid skill in its own right.

What a study on beginners shows: faster right away, but less retained

A study published in 2026 at ICER, a leading conference on computing education research, compared 22 beginners writing Python during 20-minute sessions, once with a human partner, once with GitHub Copilot. In the moment, results with Copilot were better: significantly higher performance, lower mental workload. But the emotional experience was less positive than with a human partner, and, more importantly, when the same participants were retested a week later without any help, the drop in performance was larger for those who had used the AI. The authors explicitly recommend that educators keep a place for human pair programming alongside AI, not instead of it.

In other words: what you produce quickly with AI help doesn’t necessarily stick in your memory as well as what you had to build through real effort.

What happens in your brain when you hand everything over

In June 2025, a team at the MIT Media Lab published a study that followed 54 participants (18 of them through a fourth session), split into three groups for writing essays: with no tool at all, with a search engine, or with an AI assistant, all while wearing a headset that measured brain electrical activity (EEG). The result: the no-tool group showed the strongest, most widely distributed brain activity; the AI group showed the weakest connectivity of the three. The researchers proposed the term “cognitive debt” for this gap. One notable detail: when participants used to relying on the AI were then asked to write with no help at all, their brain activity stayed lower than that of the group that had never used the AI, as if the habit left a trace afterward.

These three studies don’t measure exactly the same thing (perceived productivity, one-week retention, brain activity), and none of them says you should stop using AI: the developers in the METR study were likely still more effective on other tasks the study didn’t measure, despite the slowdown, and the beginners in the 2026 ICER study still did better on their immediate task with Copilot. What they show together is more precise than “AI is bad”: what gets produced quickly with AI isn’t automatically understood, retained, or mastered, and the feeling of having improved isn’t proof that you actually did.

What you still need to know how to do yourself

• Read code you didn’t write. GitHub Copilot’s own official documentation puts it plainly: always validate what’s suggested, and understand it before you use it. If you can’t read what the AI just produced, you can’t judge whether it’s actually correct.
• Debug. Understand an error message and trace it back to its cause, rather than pasting it into the AI and accepting the first fix offered without knowing what it actually changed.
• Test. Know how to write, or at least read and understand, a test that proves a feature actually works, rather than relying on the impression that “it looks fine.”
• Version your work. Use Git to understand what changed, when, and why, rather than losing track of successive versions.
• Understand the fundamentals. How data is structured, why a program is slow: these fundamentals are exactly what let you spot that an AI’s answer doesn’t hold up, even when it’s stated with total confidence.

None of these five skills gets learned by watching AI do the work for you: they get learned by practicing them, including when that’s slower, exactly what the 2026 ICER study on pair programming suggests.

Use AI as a tutor, not as a pen that writes for you

The good news is that none of these studies conclude you should avoid AI. What they show instead is that there’s a real difference between asking it for a ready-made answer and asking it to explain something you don’t understand. Ask it to justify its choice, to let you guess at the bug before revealing it, or to quiz you on what you just wrote instead of writing it for you. And keep some moments, as the 2026 ICER study suggests, where you solve a problem with no help at all: that’s the kind of effort that actually leaves a trace a week later.

Your school’s rules

Using AI to learn to code is one thing; using it on graded work is another, and it’s handled differently from one institution to the next. Before using it for a graded assignment or project, check your own institution’s AI usage rules: what’s tolerated, what needs to be disclosed, and what’s off-limits varies a lot from one school to another.

Key takeaways

• A July 2025 METR study, run with 16 experienced developers, measured that AI slowed them down by 19%, even though they themselves believed it had saved them time.
• A 2026 ICER study with 22 beginners showed better immediate performance with AI, but a sharper drop in retention a week later.
• An MIT Media Lab study (June 2025, 54 participants, EEG measurements) measured weaker brain connectivity among AI users, an effect that persisted even once the AI was taken away.
• Reading code, debugging, testing, versioning, and understanding the fundamentals remain skills AI cannot exercise on your behalf.
• For graded work, always check your institution’s AI usage rules before using it.

To practice reading and understanding an error message before handing it to an AI, the lire-une-erreur skill offers a step-by-step method.