What Vibecoding Can't Teach You
“You don’t need to learn to code anymore. AI can do everything.”
You have heard it. Maybe from a tech CEO on an earnings call, maybe from a LinkedIn influencer who shipped a to-do app last weekend, maybe from a friend who asked why you are still “typing code by hand like it’s 2019.”
Vibecoding, the practice of describing what you want, accepting whatever the AI produces, and never really reading it, has become a cultural moment. Some of it is genuinely exciting. People who never could have built software are building things. But underneath the excitement is something I think we are not talking about honestly: it creates a powerful illusion of competence, and it is quietly making a lot of experienced developers miserable.
Dunning-Kruger on autopilot
In 1999, psychologists David Dunning and Justin Kruger described a familiar pattern: people with the least skill in an area tend to overestimate their ability the most. Part of the reason is cruel and simple. The knowledge you need to do something well is the same knowledge you need to notice you are doing it badly.
Vibecoding supercharges that effect. Before AI, a beginner who overestimated themselves hit a wall fast. The code did not compile, the page did not load, reality delivered feedback. Now the AI clears the first wall for you. The app runs. The demo works. And because it works, it feels like mastery.
What you cannot see is everything underneath. The SQL injection in the login form. The API key committed to a public repo. The state bug that only appears with two users. The 4,000-line file nobody, including the AI, can safely change anymore. An experienced developer spots these because they have been burned by them. A vibecoder does not know they exist, so they do not know to ask.
The result is a new kind of confident beginner: someone who has shipped something, genuinely believes they are a pro, and has no way of knowing what they do not know.
It’s not just beginners: our sense of speed is unreliable too
The illusion is not limited to people who cannot code. In 2025, the research nonprofit METR ran a randomised trial with 16 experienced open-source developers working on their own large projects. With AI tools, they took 19% longer to finish tasks. The striking part: afterwards, they still believed AI had made them about 20% faster.
To be fair, METR has since said those numbers are out of date. Its early-2026 follow-up suggests newer tools probably do speed many developers up, though selection effects made the size hard to measure. But the perception gap is the finding that matters here. Even experts cannot reliably feel whether a tool is helping them. “It feels productive” is not evidence.
Then there is learning. In January 2026, Anthropic, which makes one of the most popular coding assistants, published a randomised trial of 52 mostly junior developers learning a new Python library. The group that used AI scored 17% lower on a quiz about concepts they had used just minutes earlier, with the biggest gap in debugging. They were not meaningfully faster either. The developers who kept their understanding were the ones who used AI to ask questions and explain concepts, not to write the code for them.
That is the uncomfortable pattern. The more you hand over, the less you keep.
A lot of us code because we love it
Here is what the “nobody needs to code anymore” crowd misses: for many developers, writing code was never just a means to an end.
We got into this because we loved it. The puzzle of a tricky bug. The quiet satisfaction of an elegant function. That moment at 1 a.m. when the thing you have been fighting for three hours finally clicks, and you understand not just that it works but why. Programming is a craft, the way woodworking or cooking or music is a craft. Nobody tells a carpenter they no longer need to learn joinery because IKEA exists.
When the job becomes writing prompts, reviewing machine output and babysitting agents, the part many of us loved gets cut out. What is left can look a lot like quality assurance for a very fast, very confident intern who never learns from their mistakes.
The FOMO trap
And yet we keep using it. Not always because we want to, but because we are scared not to.
Every week brings a new model, a new agent, a new “10x” workflow, and a new post implying that anyone not using it will be unemployable by Christmas. Managers ask why the ticket took two days when “AI could do it in ten minutes.” So we adopt tools out of fear, not because they make the work better. Even METR noticed the shift: by late 2025, a growing number of developers refused to join its study because they did not want to work without AI for half their tasks.
The cost is real, and it compounds:
- Skills atrophy. Abilities you do not use fade. Stop writing tricky logic yourself and, a year later, you notice you reach for the prompt box before you even try.
- Debugging gets harder. Reading someone else’s code is harder than writing your own, and AI-generated code is always someone else’s code.
- Flow disappears. Deep focus gets replaced by a loop of prompt, wait, review, correct, re-prompt. It is exhausting in a way that writing code rarely was.
- Joy drains out. The work gets faster, maybe, but it stops feeling like yours.
We are, many of us, actively using tools that make us less happy and less skilled, because we are afraid of being left behind. That is not progress. That is a treadmill.
The gold rush where everyone has a shovel
There is another fantasy riding along with vibecoding: that AI is your ticket to building the next big app and becoming a millionaire overnight.
But think about what AI actually changed. It did not give you a secret advantage. It gave everyone the same one. The same models, the same agents, the same prompts are available to anyone with a laptop and a subscription. If you can vibecode a habit tracker in a weekend, so can the next person, and so can the thousands of people who already did.
When anyone can build something, the thing itself stops being scarce, and scarcity is what people pay for. Why would someone pay you for an app they could describe to the same AI and get in an afternoon? The moat is gone. Building software used to be the hard part. Now it is the easy part, which means it is no longer where the value is.
Honestly, that should be a relief. You are not falling behind in a race, because there is no race to win. Nobody is building something with AI that you could not also build. You can relax.
So instead of trying to sell people something they could make themselves, point AI at your own life. Use it as an assistant and a coworker. Let it automate the tedious parts of your business, draft the emails you dread, organise your finances, help you learn something new, or build the small tools only you needed. The real payoff is not a product to sell. It is a better life and a better business of your own.
Use the tool. Don’t become it.
I am not arguing that AI is useless, or that nobody should use it. It is a genuinely powerful tool, and the research above suggests the problem is less the tool than how we use it. People who use AI to explain, question and teach keep their skills. People who use it to avoid thinking lose them.
So here is my case for a saner approach:
- Learn to code anyway. You cannot judge output you do not understand. The better AI gets, the more valuable real understanding becomes, not less.
- Write the hard parts yourself. Let AI handle boilerplate if you like, but keep your hands on the logic that matters.
- Use it as a tutor, not a ghostwriter. Ask why. Ask for the trade-offs. Make it explain itself.
- Protect the joy. If you love coding, carve out time to just code. That is not inefficiency. That is how you stay good, and how you stay sane.
Vibecoding can build you an app. It cannot make you a developer. And it definitely cannot make you love this work the way many of us still do.