What is vibe coding and why does it get so much hate?

Vibe coding, defined

Vibe coding is a form of software development that uses a large language model to generate some or all of the code for a project. You describe what you want in plain language; the model writes the implementation. Skip this if you already hand every non-trivial function to an LLM and only review the diff — that's the practice, whatever you call it.

The term came from Andrej Karpathy, the AI researcher best known for leading Tesla's Autopilot Vision program. In February 2025 he posted on X: "There's a new kind of coding I call 'vibe coding,' where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good."

The framing is the point. Traditional AI-assisted coding still treats the model as a subordinate — it autocompletes, you decide. Vibe coding flips that: the model produces, you prompt and accept. The word "forget" is doing real work in Karpathy's definition. It's not ignorance of programming; it's deliberate non-attention to the artifact itself.

Why it gets so much hate

Skip this if you want a cheerleading post. The criticism is the interesting part, and much of it is justified.

Proponents argue LLMs democratize software development, letting anyone build their own apps without a CS degree. That is a real change — the barrier to entry has collapsed. The people who celebrate that are not wrong about what it enables.

Critics are worried about what you end up with. They argue vibe coding is producing vulnerable software, and that extensive use can make a codebase more difficult to maintain. You cannot review what you do not understand, so the bugs ship and the fixes hurt.

The resentment runs deeper than code quality, though. Coding was one of the few reliable ladders into the middle class, and companies are now hiring fewer junior coders than ever before. A tool that devalues the craft hits people where their paycheck lives. My honest read: a lot of the hate is grief for a path that is already closing, aimed at the loudest new symbol of it.

The uncomfortable part is that both sides keep being right at the same time. That is why the argument stays loud.

What the security research says

AI-generated code ships with real vulnerabilities, and Georgia Tech has the receipts. In April 2026, researchers from the School of Cybersecurity and Privacy scanned 43,000 security advisories and traced 74 vulnerabilities directly back to AI-generated code. Fourteen of those were severe enough to be classified critical.

The scan was narrow by design. Those 43,000 advisories only covered a three-month window at the start of the year. And the team could only trace code that had been properly disclosed as LLM-generated — which almost nobody does. Their estimate puts the real count at five to 10 times higher than the 74 they confirmed.

My read: the 74 is the floor, not the finding. The headline number matters less than what it implies about disclosure — most AI-assisted code slips into production without anyone flagging it as such.

Vibe coding by the numbers: who actually uses it

Adoption is not the problem. A survey of 1,100 professional programmers who have tried AI tools found that 72 percent use them every day, and around 42 percent of their codebase is already AI-generated or AI-assisted. The same group expects AI-written code to cross the halfway mark next year. Stack Overflow's 2025 survey put daily AI-tool usage slightly lower, at 47.1 percent, but the direction is unmistakable.

What professionals actually do with these tools is far more conservative than the hype suggests. As of mid-2025, most were using AI for autocomplete, code review, and chatting through roadblocks — the kind of work that speeds up an existing workflow without handing over the design decisions. This is AI-assisted development, not vibe coding.

That distinction shows up clearly in the Stack Overflow data. 72 percent of respondents said vibe coding was not part of their workflow, and another 5 percent said it was emphatically not how they do their job. You can adopt every AI tool available and still reject the idea of letting the model drive.

Skip the rest of this debate if you're looking for a clean adoption-versus-backlash story — the numbers don't give you one. What they show is a workforce that has already absorbed AI into daily practice while drawing a hard line at surrendering control.

AI-assisted vs. vibe coding: the difference that matters

AI-assisted means you stay in charge of the code. It's the difference between an editor and a ghostwriter. You write the commit, you review the diff, and you own the bug when it ships. As of mid-2025, most professionals were using AI tools for autocomplete, code review and chatting through roadblocks — the software equivalent of a second pair of eyes, not a replacement for yours.

Vibe coding inverts that relationship. You describe what you want in plain language, the model writes everything, and you ship whatever it produces. You're not reviewing logic anymore — you're prompting, waiting, and hoping. When it works, it feels like sorcery. When it doesn't, you're debugging code you've never actually read.

Skip this if you don't care who wrote the code. But if an outage at 2 a.m. lands in your lap, the distinction is the only thing standing between fixing it and reading it for the first time. That's the difference that matters, and it's one nobody in this space frames clearly enough.

The cost of vibe coding for new developers

Companies are hiring fewer junior coders than ever before. Vibe coding accelerates that trend — if a senior engineer can describe a feature to an AI and get working code, the business case for hiring someone to write boilerplate collapses. You don't need to take that on faith: coding used to be one of the few reliable paths into the middle class, and that's no longer the case.

The senior engineers aren't the ones losing sleep. They gain leverage — one person with tooling does what a small team did in 2023. It's the entry-level pipeline that breaks first. The roles that used to teach juniors how production systems work are exactly the roles AI now absorbs. Skip this section if you're mid-career; this particular problem isn't yours yet. My honest read: the industry hasn't figured out how to train its next generation without those junior seats, and nobody's talking about that gap.