What Happens to Vibe Coders? What the Research Actually Says

#ai #beginners #webdev #programming

Vibe coding went from a joke tweet to a daily habit in under two years. You describe what you want, an AI writes the code, and you ship. It is fast, it is fun, and it has let a lot of people build things they could not build before.

But what actually happens to developers who work this way? Researchers have started to measure it. I went through the main studies and surveys from 2025 and 2026 and put the results in plain language, with links so you can check everything yourself.

This is not an argument for or against AI. It is a look at what the data says, and what it means for how you can get the most out of these tools.

TL;DR

Most developers now use AI tools, but only a minority say they vibe code in their professional work.
In one randomized trial, developers who let AI do everything learned less and were worst at debugging.
Developers who used AI to ask questions and understand kept their learning almost intact.
Experienced developers can feel faster with AI while measuring slower. Feeling and measuring are not the same thing.
AI-written code often works but is not always secure, so review and testing still matter.
The research is early and the studies are small. Treat it as a useful signal, not a verdict.

What is vibe coding?

The term was coined by Andrej Karpathy in a post on 2 February 2025. He described a style where you give in to the flow, talk to the AI, and stop looking at the code. In his words: "I 'Accept All' always, I don't read the diffs anymore."

Developer Simon Willison gives a useful definition: vibe coding is building software with an LLM without reviewing the code it writes. If you review it, test it and can explain it, that is regular software development with a powerful assistant.

So the dividing line is not "do you use AI?" It is "do you understand what the AI gave you?"

How many developers actually vibe code?

Less than the headlines suggest. The 2025 Stack Overflow Developer Survey shows a gap between using AI and vibe coding:

Question	Result
Using or planning to use AI tools	84% (up from 76% the year before)
Professional developers using AI daily	51%
Say vibe coding is not part of their professional work	72%
Distrust the accuracy of AI output	46%
Highly trust the accuracy of AI output	3.1%

AI is now normal. Blind trust is not.

Does vibe coding hurt your coding skills?

This is the question that Anthropic tested in a randomized controlled trial. Here is the setup in simple terms:

52 mostly junior developers learned a Python library they had never used (Trio).
Half could use an AI assistant. Half could not.
Afterwards, everyone took a quiz on what they had learned.
The results:
The AI group scored about 50%. The no-AI group scored about 67%.
The AI group finished only about two minutes faster, which was not statistically significant.
The biggest gap was in debugging, which means spotting that code is wrong and understanding why.
That sounds bad for AI, but the study found something more interesting. How people used the AI changed everything:

How they used AI	Quiz score
Generated code, then asked follow-up questions to understand it	65% or higher
Asked for code and explanations together	65% or higher
Asked only concept questions and fixed errors themselves	65% or higher
Let AI write everything	Below 40%
Started by asking, then slowly handed over everything	Below 40%
Let AI debug instead of understanding the bug	Below 40%

The tool was the same. The habits were different. Developers who stayed mentally active kept their learning. Developers who fully delegated did not.

A note on fairness: Anthropic makes Claude, so this is a study from an AI company, and it measured learning a new library, not general programming ability.

What happens in the brain?

Nobody has scanned the brains of vibe coders yet, so this part combines two lines of research.

1. What coding asks of your brain. MIT fMRI studies found that understanding code mostly uses the brain's problem-solving and working-memory network, not the language centers that handle normal sentences. A follow-up paper found that this same network is best at picking up what a program will actually do when it runs.

2. What happens when AI goes first. In the MIT Media Lab study Your Brain on ChatGPT, 54 people wrote essays while wearing EEG caps. People using ChatGPT showed the weakest brain connectivity, struggled to quote their own writing, and felt less ownership of it. The researchers call this "cognitive debt": you save effort now and pay for it later.

There was a hopeful twist. People who wrote on their own first and used AI afterwards did better than people who started with AI. Order mattered.

Two honest caveats: this was a small preprint about essays, not code, and my comparison below is my own reading, not a measured result.

	Developer who writes the code	Vibe coder
Where the effort goes	Tracing logic, holding behavior in mind	Describing goals, checking the result
Mental model	Architecture and why each part exists	Prompts and what the app seemed to do
When it breaks	Forms a hypothesis and tests it	Pastes the error back and tries again

The effort that writing code demands is also the effort that builds skill. When AI takes it over, the skill-building can quietly stop, unless you add it back on purpose.

Does AI actually make you faster?

Usually it feels like it does. The numbers are more mixed.

METR ran a randomized trial with 16 experienced open-source developers working on 246 real tasks in their own projects. With AI allowed, they took 19% longer. Before starting, they expected AI to make them 24% faster. Afterwards, they still believed it had made them 20% faster.

The authors are careful about this result. It covers experienced people, large familiar codebases and early-2025 tools. It does not mean AI slows everyone down. It does show that how fast something feels can differ from how fast it is.

Is AI-generated code secure?

It often works. Whether it is safe is a separate question.

Veracode tested over 100 AI models on 80 coding tasks. In 45% of cases, the code introduced a vulnerability from the OWASP Top 10, the list of the most serious web security risks. Newer models were better at producing working code but not noticeably better at producing secure code.

Developers sense this. In Sonar's 2026 survey of 1,149 professional developers, 96% said they do not fully trust AI-generated code to be functionally correct. Juniors were more likely than seniors to say AI code "looks correct but isn't reliable" (66% vs 48%), which makes sense: spotting a subtle problem is easier when you have seen many of them.

And in the Stack Overflow survey, 66% named "almost right, but not quite" as their biggest frustration with AI.

So how do you get the best of both?

Putting the studies together, a few habits keep showing up. None of them mean using AI less. They mean using it more deliberately.

Ask "why" after you get code. The best-scoring group in the Anthropic trial generated code and then asked follow-up questions.
Think first on hard problems. The MIT results suggest starting with your own idea and then bringing in AI.
Form a guess before you paste an error. Handing all debugging to AI was one of the lowest-scoring patterns.
Know your architecture. You should be able to sketch what the main parts are and how data moves between them.
Test and scan. Write tests for the paths that matter, and run a security scanner on AI-written code.
Practice without AI sometimes. Small projects from scratch keep your fundamentals sharp.
Be able to take over. If the AI were unavailable tomorrow, you should still be able to continue.

What the research can't tell us yet

Most studies are small. Anthropic had 52 people, METR had 16, and MIT had 54.
Some tested learning something new, not everyday work on code you already know.
Tools improve quickly, so results from early 2025 may not hold for today's tools.
No one has studied the long-term effects of years of vibe coding.

Final thoughts

Vibe coding is a real and useful way to build, especially for prototypes, side projects and getting past a blank page. The research does not say "stop." It says the outcome depends on what you do while the AI is working: whether you stay curious about the code or let it pass by.

Use AI to move faster. Use it to explore. And keep understanding what you ship, because that is the part that makes you a more capable developer over time.

What about you? Do you read every diff the AI gives you, or accept and move on? I would like to hear how you work in the comments.