Vibe Coding Is Eating Software Development. Not Everyone Is Happy.

TL;DR: Vibe coding, accepting AI-generated code without reading it, went from a Karpathy tweet to Collins Word of the Year to an MIT breakthrough technology in under a year. The data says both camps are right: 84% of developers use AI tools and individual output is up, while duplicated code quadrupled, vibe-coded apps ship 2.74x more security vulnerabilities, and satisfaction is falling. My position after using AI daily on client work: the tools are real leverage, but only when a human stays accountable for every line that ships.

Andrej Karpathy coined vibe coding in a post on X in February 2025. The post got 4.5 million views. By March, Merriam-Webster added it as a slang entry. By November, Collins Dictionary named it Word of the Year. By January 2026, MIT Technology Review listed generative coding as a 2026 Breakthrough Technology. By February 2026, Karpathy himself called the term "passe" and introduced "agentic engineering" as the next evolution.

That's a fast arc for a concept that started as a casual post about accepting AI output without reading the diffs. Here's my take: vibe coding is real, it's useful, and the people most upset about it are largely upset for the wrong reasons.

What is vibe coding, actually?

The original definition from Karpathy is blunt: you describe what you want, the AI generates code, and you "forget that the code even exists." You accept output without reviewing it, nudge it with follow-up prompts, and ship when it works.

Simon Willison drew a line that matters more than the hype. If you have reviewed, tested, and understood the code, that is not vibe coding. That's using an LLM as a typing assistant. The distinction is about accountability, not about whether AI touched the code.

That's exactly how I work. I review every important piece, test it, and understand the architecture before anything goes near production. When I built FlowMate, an email-management SaaS I've since shelved, every line of AI-assisted code went through manual review. The AI accelerated the typing. The engineering decisions were mine, and so were the mistakes.

But plenty of people use the Karpathy definition literally. Non-coders building apps. Founders shipping MVPs without a technical hire. Solo developers building internal tools faster than any team could. That's the part making senior engineers nervous.

What do the numbers say about AI coding in 2026?

The Stack Overflow 2025 Developer Survey found 84% of developers now use or plan to use AI coding tools, up from 76% in 2024 and 70% in 2023. Over half use them daily. GitHub Copilot crossed 20 million users, and GitHub's own research has long put Copilot at nearly half the code in files where it's enabled.

The claims from tech leadership keep escalating. Satya Nadella said 20 to 30% of Microsoft's code is now written by software. Dario Amodei at Anthropic has claimed publicly that the large majority of their code comes from Claude. At Y Combinator's Winter 2025 demo day, Garry Tan revealed that 25% of startups in the batch had codebases that were almost entirely AI-generated.

The part that gets less attention: positive sentiment toward AI coding tools dropped. The same survey showed satisfaction fell from over 70% in previous years to 60% in 2025, and 66% of developers reported frustration with AI solutions that are "almost right, but not quite." Adoption is real. Satisfaction is not keeping up.

Is AI-generated code worse? The studies, side by side

The concerns are not imaginary. The research through early 2026:

The Wiz pattern was always the same: client-side authentication, hardcoded API keys, unprotected database access. And DORA's nuance matters: separate engineering-telemetry research from Faros AI measured individuals completing about 21% more tasks and merging 98% more pull requests, yet organizational delivery metrics stayed flat. Faster typing does not automatically mean faster shipping.

Someone who vibe-coded a customer-facing app with no understanding of the security model has created a liability, not a product. As more of these hit production, someone has to clean them up.

Why are open source maintainers banning AI code?

One consequence caught the industry off guard: AI-generated pull requests started flooding popular repositories. Daniel Stenberg, maintainer of curl, wound down his bug bounty after the share of genuine vulnerabilities in submissions fell to about 5%. Mitchell Hashimoto banned AI-generated code from Ghostty entirely. Steve Ruiz closed all external pull requests to tldraw. GitHub started considering restrictions on pull requests to help maintainers cope.

This is an active crisis for the people who keep critical infrastructure running, and it's a direct result of vibe coding applied where it doesn't belong: existing, complex systems that require deep understanding before you touch them.

The criticism is also missing the point

The loudest complaints are framed as "vibe coding is bad for software." What I keep reading between the lines is something else: "people who don't know what I know are now building things." That's not a technical argument. That's a guild protecting its territory.

I build websites and automation systems for local Polish businesses. Small restaurants, nail salons, barbers. None of them need a team of senior engineers. They need a working website with a contact form and decent SEO. I use AI-assisted coding to build those faster and cheaper. That's not a crisis. That's the market working.

The crisis is real for companies that vibe-coded their way into 50,000 lines of AI-generated spaghetti and now need to add a feature. The solution is not "ban AI from development." It's "understand what you are building." The dead internet problem applies to code too: quality and accountability still win.

What actually matters in 2026

I'll take a stance: the developers scared of vibe coding are the ones whose value was knowing syntax and APIs. That was always a fragile position. The developers who are fine are the ones whose value is knowing what to build, why, and whether what got built is correct.

Vibe coding raises the floor. Anyone can produce working code for simple problems, and that's genuinely good. The ceiling (understanding systems, making architectural decisions, spotting the security hole the AI missed) has not moved. If anything it got more valuable, because someone has to supervise all these AI-generated applications.

The Stanford Digital Economy Lab found that employment among software developers aged 22 to 25 fell nearly 20% between 2022 and 2025, while developers over 26 saw stable or growing employment. The entry-level "write code from tutorials" job is disappearing. The "understand systems and make decisions" job is not.

I use Claude on almost every project now, from AI integrations to the automation workflows I built to find businesses without websites. I prompt, I review, I modify. I don't fully give in to the vibes. But I also don't pretend I hand-write everything to preserve a purity that stopped mattering two years ago.

The middle position

The developers who will struggle are not the ones who adopt AI coding. They're the ones who refuse it, and the ones who adopt it without judgment. The middle position, where you use the tools and stay responsible for the output, is where the actual work happens.

Vibe coding is eating software development. Some of that is messy and some will cause problems. But the core shift, that describing what you want in plain language is now a valid way to build software, is here and not reversing. The question was never "should developers use AI?" It was always "how do you use it without creating a mess?" The answer is what it has always been in engineering: understand what you're building, test it, and take responsibility for the result.

This article is also available on Medium.