The Vibe Coders

A product manager at a Series B startup builds an internal analytics dashboard in an afternoon — three APIs, authentication, clean UI, deployed to production. The engineering team finds out when someone sends a Slack message asking who maintains it. Their reaction: impressed, then terrified. Then, quietly, a little displaced.

This is not a hypothetical. Variants of this story are now common enough to have a name, a Wikipedia article, and a Collins Dictionary Word of the Year entry. The name is "vibe coding." The person who named it is already trying to move past it.

The Tweet That Named the Moment

On February 6, 2025, Andrej Karpathy — former director of AI at Tesla, early OpenAI researcher, one of the most credentialed AI researchers alive — posted a description of how he'd been building lately. "There's a new kind of coding I call 'vibe coding,'" he wrote, "where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." He described using Cursor with Claude Sonnet, barely touching the keyboard, accepting all diffs without reviewing, copy-pasting error messages back to the model without reading them. "It's not really coding," he acknowledged. "I just see stuff, say stuff, run stuff, and it works."¹

The tweet got 4.5 million views. Developers immediately split into two camps. One camp saw the democratization of software creation, the logical endpoint of decades of abstraction. The other saw an admission that professional skills didn't matter anymore, delivered cheerfully by someone whose professional skills were not at risk.

By late 2025, Collins Dictionary named "vibe coding" Word of the Year. Merriam-Webster countered with "slop." The editorial choice was pointed.

On the one-year anniversary, Karpathy posted a retrospective. He noted that the original tweet was "a shower of thoughts throwaway." He introduced "agentic engineering" as the term he actually wanted to use — "agentic because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents… engineering to emphasize that there is an art & science and expertise to it."¹ He was trying to walk back the casual framing. It didn't work. The genie was already out of the bottle, and the genie wasn't interested in being called an engineer.

The Numbers

By early 2026, 81% of CIOs reported that employees were creating AI agents and applications faster than IT could govern them. More than half had discovered unsanctioned "shadow AI" in their organizations.⁶ The people building these tools were not primarily developers — they were marketers, designers, product managers, solopreneurs, and founders who'd decided to bypass the engineering queue entirely. Meanwhile, 63% of developers who did use AI coding tools reported spending more time debugging AI-generated code than writing the original code themselves would have taken.²

21% of Y Combinator's Winter 2025 cohort had codebases that were 91%+ AI-generated. 46% of all new code written globally is now AI-generated. 25% of Google's internal code is AI-assisted. The AI coding tool market is projected at $4.7 billion in 2026, growing to $12.3 billion by 2027.²

And yet: 72% of professional developers say vibe coding plays no role in their work. Only 15% have adopted it to any degree. Just 0.4% describe themselves as enthusiastic practitioners.³

The people building unsanctioned AI applications are largely not software developers. The people most threatened by vibe coding are mostly not doing it. The two populations are looking at each other from across a gap that is widening by the quarter.

What It Actually Looks Like Under the Hood

In December 2025, CodeRabbit analyzed AI-co-authored code at scale. The findings were not reassuring: AI-generated code has 1.7x more major issues than human-written code, 75% more logic errors, 75% more misconfigurations, and 2.74x more security vulnerabilities.⁴ These are numbers for professional developers using AI as a tool. For non-developers using AI as a replacement for understanding, the numbers are worse.

In May 2025, a security researcher audited apps built on Lovable, an AI app-building platform popular with non-technical founders. He found that 170 of 1,645 apps — 10.3% — had critical row-level security flaws in their Supabase configurations. These weren't hobby projects. They were handling real user data.⁴

Bas Nijholt, a developer who documented his own vibe coding experience in detail, built a working prototype in three hours. Cleanup took fifteen. The AI-generated code contained "redundant checks for impossible conditions, irrelevant error handling, execution paths that would never trigger." His term for what he'd been given: a "technical prison" — superficially functional, internally chaotic, expensive to escape.⁵

Amazon mandated 80% weekly usage of its AI coding assistant Kiro across engineering teams in early 2026. The result was a six-hour outage affecting checkout, login, and product pricing — roughly 6.3 million lost orders. Root cause: engineers accepted AI-generated code at high rates without catching subtle flaws. The fix — mandatory senior engineer approval for all AI-assisted production deployments — created a review bottleneck that defeated most of the productivity gain.⁶

The Hashnode State of Vibe Coding 2026 survey found that code churn has increased 41%, code duplication has increased 4x, and refactoring has collapsed from 25% of development time in 2021 to under 10% in 2024. 41% of developers admit pushing AI-generated code to production without full review.²

That's professional developers. Who reviews the vibe coder's production push?

The Professional Reaction

Developers' favorability toward AI tools dropped from 77% in 2023 to 60% in 2026. Trust in AI code accuracy dropped from 43% to 33% over the same period. 75% of developers, when uncertain, still prefer consulting a colleague over asking an AI.³

The METR study published in July 2025 produced the most damaging dataset. Sixteen experienced open-source developers were given real tasks, randomized to use AI or not. The result: developers with AI tools were 19% slower. Before the study, they'd predicted AI would speed them up by 24%. After experiencing the slowdown, they still believed they'd been 20% faster. The subjective experience of productivity and the objective measure of productivity had completely decoupled.³

The developers who report feeling productive while measurably producing lower-quality code aren't lying. The METR subjects believed they were 20% faster even after the study showed they were 19% slower. They're experiencing something real. They're just not experiencing what they think they're experiencing.

The professional-versus-vibe-coder split runs deep. Most developers have settled on a rhetorical distinction between "AI-assisted development" (what they do, legitimate) and "vibe coding" (what non-developers do, not). The tools are identical. The framing is everything. It's how a profession maintains its identity while the ground shifts under it.

We Have Seen This Movie Before

In 1985, the desktop publishing revolution arrived. Before it, all typesetting was done by roughly 4,000 specialist firms across North America. After it, anyone with a Mac and PageMaker could produce print-ready pages. Specialists were "rapidly replaced by clerical workers or professionals in other disciplines whose chief qualification was operating a computer."⁷ By 2000, the typesetting industry had largely ceased to exist as a standalone profession.

But graphic design didn't die. It stratified. The low end got absorbed by non-specialists. The high end became more valuable, more conceptual, more focused on craft that machines couldn't replicate. The profession transformed rather than disappeared.

WordPress went live in 2003 with the explicit mission to "democratize publishing." It now powers 43% of the internet. It created an entirely new tier of "web developer" — people who configure themes and plugins without touching code. Every professional web developer predicted it would destroy the profession. Instead it moved the profession upmarket. The WordPress ecosystem created more developer jobs than it eliminated, just different ones.⁷

The pattern from both waves: moral panic, market segmentation, profession redefinition. Low-skill work gets absorbed by the new tool. Skilled practitioners move to work the tool can't handle. The profession shrinks at the bottom and expands at the top.

The question for software development is whether AI coding tools are just the latest wave in this pattern, or whether they represent something qualitatively different. The previous democratization tools — desktop publishing, WordPress, no-code platforms — all hit a ceiling. They handled simple work and bounced off complexity. AI coding tools don't obviously have that ceiling. They handle increasing complexity, iterate on failures, and improve with every model generation. This is not necessarily the same wave.

The Handoff Problem

Here's what the success stories don't mention: who maintains it.

Pieter Levels built an MMO flight simulator in 30 minutes that scaled to hundreds of thousands of users and generates $50,000+ per month. This is a real achievement. Levels is also an experienced developer with deep architectural understanding. He is vibe coding on top of expertise. He knows when the vibe is wrong.⁸

The product manager who built the analytics dashboard is not Pieter Levels. The dashboard is in production. It pulls from three APIs. Three months from now, one of those APIs will change. The dashboard will break. The product manager will have moved on to the next project — or left the company. A developer will be handed a codebase with no architecture, no tests, no comments, and no author who can explain the intent. They will spend more time decoding it than they would have spent building it from scratch. This is not hypothetical. It is a pattern that has a name now: the ghost codebase.

The vibe-coding disillusionment trend documented by App Builder Guides in early 2026 shows the arc: non-developers hit the 80/20 wall. The first 80% of a project — the thing you can describe in plain English — comes fast. The last 20% — edge cases, integrations, production hardening, security, maintenance — requires exactly the skills vibe coding promised weren't necessary. At that wall, builders either abandon the project, return to structured no-code platforms, or hand the mess to a professional developer who didn't build it and can't fully understand it.⁵

The Genuine Value

It would be dishonest to tell only the disaster stories.

Christine Hudson, a researcher with a master's in machine learning who hadn't coded in fifteen years, used AI coding tools to build a working Google Calendar export utility — the only person in her cohort to complete the task. She described it as producing "experiences of pure joy and effortless creation."⁸ She's not a vibe coder in the pejorative sense. She's a domain expert who temporarily lost access to technical implementation and got it back.

There is a class of software that vibe coding creates that would not otherwise exist: small internal tools solving specific problems for specific teams, personal projects that scratch particular itches, niche solutions for niche industries that can't justify hiring a developer. For these uses, some vibe-coded software with security debt is better than no software at all.

The honest accounting is that vibe coding is a tool with a legitimate use case and a common misuse. The legitimate use case is: domain experts building small-scope tools in domains they understand, with an awareness of the tool's limitations. The common misuse is: anyone building anything production-grade without that awareness.

The problem is that nothing about vibe coding communicates the distinction between these two uses. The tool doesn't know which one you're doing. And the vibe coder, almost by definition, doesn't know what they don't know.

The Profession's Choice

The profession of software development now faces a version of the question graphic designers faced in 1985 and web developers faced in 2003. It can respond with gatekeeping — insisting on credentials, dismissing vibe coders as not real developers, protecting its identity through exclusion. Or it can respond by moving upmarket — becoming the people who maintain what vibe coders build, who architect systems complex enough that AI can't prompt them into existence, who translate between the business problems that vibe coders understand and the engineering constraints they don't.

The gatekeeping response is instinctive and understandable and will not work. The vibe coders are already in production. They are not going to stop because developers find them illegitimate.

81% of CIOs say employees are creating AI agents and applications faster than IT can govern them. More than half have discovered unsanctioned "shadow AI" in their organizations.⁶ The governance frameworks are running 12 to 18 months behind adoption. The vibe-coded software is already there, already running, already integrated into business processes that someone will eventually need to maintain.

Karpathy tried to walk back his own term because he understood that "vibe coding" implied something he didn't intend: that expertise was optional, that craft didn't matter, that forgetting the code exists was the goal rather than a useful temporary shortcut. He's right that the framing was wrong. He's also right that it's too late to change it.

The question for the profession is no longer whether non-developers will build production software. They are. The question is who gets paid to clean up after them — and whether that work is going to be called software engineering or something else entirely.