# Is vibe coding bad for learning to program?

Source: https://codelifter.com/forum/posts/is-vibe-coding-bad-for-learning-to-program/
Site: CodeLifter Forum, category Code provenance
Published: 2026-10-24. Updated: 2026-10-26. Replies: 2.

## Question by CodeLifter editors (staff), 2026-10-24

Vibe coding is bad for learning when you hand the whole task to the model, and much less so when you use AI to explain. In a January 2026 randomized trial of 52 developers, the AI group scored 17% lower on a comprehension quiz, but those who asked conceptual questions scored about as well as people coding by hand.

## What is vibe coding?

Vibe coding is writing software by describing what you want to an AI assistant and accepting what it produces without reading it closely. Merriam-Webster's [slang entry](https://www.merriam-webster.com/slang/vibe-coding) defines it as writing computer code "in a somewhat careless fashion, with AI assistance", where the coder does not need to understand how or why the code works. It traces the term to a post by Andrej Karpathy on X on 2 February 2025 that talks about forgetting "that the code even exists." The longer history is in [what is vibe coding, and where did the term come from?](/forum/posts/what-is-vibe-coding-and-where-did-the-term-come-from/)

That definition already answers part of the question. Learning to program means learning to understand how and why code works, and the definition makes not understanding the point.

## What does the research say about learning?

The most direct study so far is Anthropic's randomized controlled trial, [How AI assistance impacts the formation of coding skills](https://www.anthropic.com/research/AI-assistance-coding-skills), published on 29 January 2026 (paper: arXiv 2601.20245). A randomized controlled trial assigns people to groups by chance so that differences in results can be put down to the treatment, here access to an AI assistant.

- **Who:** 52 mostly junior software engineers who used Python weekly but had not used Trio, an asynchronous Python library.
- **Task:** build two features with Trio, then take a quiz on the concepts they had just used.
- **Result:** the AI group averaged 50% on the quiz against 67% for the hand-coding group. The authors describe it as nearly two letter grades. The largest gap was on debugging questions.
- **Speed:** the AI group finished about two minutes faster on average, which was not statistically significant.

The study also grouped people by how they used the assistant. This part is observational, and the authors say it does not establish cause:

| How the participant used AI | Participants | Quiz result |
| --- | --- | --- |
| AI delegation: the assistant wrote the code | 4 | Low-scoring group, average under 40% |
| Progressive reliance: a few questions, then full delegation | 4 | Low-scoring group |
| Iterative AI debugging: the assistant fixed each error | 4 | Low-scoring group, and slower |
| Generation, then comprehension questions | 2 | High-scoring group, 65% or above |
| Asking for code together with an explanation | 3 | High-scoring group |
| Conceptual questions only, writing the code yourself | 7 | High-scoring group, and second fastest overall |

Two limits are worth stating. The sample was small and the quiz came minutes after the task, so it says nothing yet about long-term skill. And the authors note in a footnote that their setup was a chat assistant in a sidebar, not an agentic tool such as Claude Code, and that they expect the effect of agentic tools on skill development "to be more pronounced."

## Does it at least make you faster?

Not reliably, and that removes the usual trade-off argument. In METR's July 2025 trial, 16 experienced open source developers working on 246 issues took 19% longer when allowed AI tools, while expecting a 24% speed-up and believing afterwards they had been about 20% faster. METR has since marked those results as out of date and published newer ones, so treat the number as a snapshot. The point that lasts is the gap between how fast it feels and how fast it was.

## What does a beginner miss when the model writes everything?

- **Security.** In the [Perry et al.](https://arxiv.org/abs/2211.03622) user study, people with an AI assistant wrote significantly less secure code and were more likely to believe it was secure.
- **Dependencies.** Models invent package names; see [why AI code imports packages that do not exist](/forum/posts/why-does-ai-generated-code-import-packages-that-do-not-exist/).
- **Debugging.** The skill with the largest gap in the Anthropic trial is the one you need most when generated code fails.

## How do you use AI without skipping the learning?

1. **Write the first attempt yourself.** Then ask the assistant to review it, not replace it.
2. **Ask conceptual questions.** "Why does this need `await`?" builds the knowledge the high-scoring group showed. "Write the function" does not.
3. **When you do ask for code, ask for the explanation too,** and read it before you paste.
4. **Debug before you ask.** Read the error, form a guess, test it. Then ask the assistant to check your reasoning.
5. **Use a learning mode where your tool has one.** In Claude Code, run `/output-style learning`. Its [output styles documentation](https://code.claude.com/docs/en/output-styles) (read 16 October 2026) says the Learning style explains its choices and leaves `TODO(human)` markers where you write the code yourself.
6. **Explain it back.** If you cannot explain a block to someone else, you have not learned it yet. A teacher may ask you to do exactly this; see [can a teacher tell if a student's code was written by AI?](/forum/posts/can-a-teacher-tell-if-a-students-code-was-written-by-ai/)

CodeLifter's desk article [Is vibe coding bad?](https://codelifter.com/provenance/is-vibe-coding-bad/) covers the same question for working developers, including copyright and security.

## Short version

- Vibe coding by definition skips understanding the code, which is the thing a learner is trying to build.
- In Anthropic's January 2026 trial, AI users scored 50% on a comprehension quiz against 67% without AI, with the biggest gap on debugging.
- People who asked conceptual questions or asked for explanations kept high scores; people who delegated did not.
- Speed gains are not guaranteed; METR's 2025 trial found experienced developers slower with AI.
- Write first, ask why, debug yourself, and use a learning mode if your tool has one.

## Reply 1 by CodeLifter editors (staff), 2026-10-24

**Should beginners use AI to learn to code at all?** Yes, if it is used as a tutor rather than a ghostwriter. The Anthropic trial is the best evidence on the difference. Its highest-scoring group, the 7 participants who only asked conceptual questions and wrote the code themselves, hit many errors but resolved them on their own, and finished faster than every other high-scoring group. The authors write that cognitive effort, even getting painfully stuck, is likely important for building mastery, and note that major AI services now offer learning modes built for this.

In Claude Code that mode is an output style. The [documentation](https://code.claude.com/docs/en/output-styles), read on 16 October 2026, says the Learning style adds short `Insight` blocks explaining its choices and, at design decisions such as error handling or a data structure, leaves a `TODO(human)` comment and waits for you to write that part. You can switch with `/output-style learning`, or set `"outputStyle": "Learning"` in `.claude/settings.local.json`; the setting value is case-sensitive, so `learning` there gives you the Default style. The docs also note that an output style is an instruction, not a guarantee, so the discipline still has to come from you.

## Reply 2 by CodeLifter editors (staff), 2026-10-26

**Does AI make programmers faster?** It depends on whether they already know the work. The two strongest studies point in different directions because they ask different questions. METR's [July 2025 trial](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) gave 16 experienced open source developers 246 issues and found they took 19% longer with AI tools allowed, despite expecting to be faster. METR has marked that page "out of date" and released results current as of early 2026, so check the newer post before quoting the figure.

Anthropic's January 2026 [skills trial](https://www.anthropic.com/research/AI-assistance-coding-skills) found no statistically significant speed gain for people learning a new library, partly because some spent up to 11 minutes, about 30% of the time, writing up to 15 queries. The same post cites Anthropic's earlier observational study, in which AI cut the time for some tasks by 80%, and reconciles the two: that work measured tasks where people already had the skills. The authors suggest AI may speed up skills you have while slowing the learning of new ones. For a student, the second half is the part that applies.

---
Cite as: CodeLifter Forum, "Is vibe coding bad for learning to program?", https://codelifter.com/forum/posts/is-vibe-coding-bad-for-learning-to-program/, accessed 2026-11-12.