Vibe Coding is dead. And that's not bad news

What happened to vibe coding in 2025–2026

February 2025. Andrej Karpathy — co-founder of OpenAI and former head of AI at Tesla — publishes a post on X describing a new approach to development: you just write a prompt, AI generates code, you iterate based on feeling without really looking inside. He called it vibe coding and added: "I even forget that code exists". Merriam-Webster immediately added the term to its "slang & trending" list.

The idea exploded. Collins Dictionary named "vibe coding" the word of 2025. By early 2026, over 80% of developers reported using or planning to use AI tools for coding. It was estimated that 41% of all new code worldwide was already being generated with AI — and this was not the ceiling: Sundar Pichai noted in a public post that at Google, this figure had reached 75%. Products that previously took months were assembled over weekends. It seemed like magic.

At the same time, alarming signals were accumulating. The Stack Overflow Developer Survey 2025 recorded a paradox: 84% of developers use or plan to use AI tools — but 46% actively distrust the accuracy of their results. The top complaint: "almost right, but not quite" — cited by 66% of respondents. The METR organization conducted a randomized controlled experiment and found a shocking result: experienced open-source developers were 19% slower with AI tools than without them — although they themselves expected a 24% acceleration and after the experiment believed they had become 20% faster. The gap between perception and reality turned out to be striking.

And then came May 2026.

Two engineers who built the core of the popular AI agent OpenClaw — Mario Zechner (creator of libGDX and the Pi framework) and Armin Ronacher — published a warning in the Wall Street Journal that resonated throughout the industry. They called the phenomenon "vibe slop" — a combination of "vibe coding" and "AI slop": the moment when thoughtless code generation devolves into chaos, leaving behind bloated, unrefactored code, crutches instead of solutions, and technical debt that hardens like concrete.

"We have infrastructure that is falling apart, and software that has become much buggier than before" — Mario Zechner, Wall Street Journal, May 2026

Ronacher, in his essay "Some Things Just Take Time", warned of "vibe slop at inference speeds" — the ability of AI to generate low-quality code at a speed that makes human oversight impossible. He also described "agent psychosis" — a state where developers become increasingly detached from the code they ship.

This was not a critique from Luddites. It was a warning from people who had themselves built a popular AI product on the same framework they were cautioning against — and who knew better than anyone where the system was breaking.

AI slop: when speed became the enemy of quality

To understand the scale of the problem, let's look at the numbers that have accumulated by mid-2026. This is not alarmism, nor the thoughts of skeptics. These are data from independent research.

Code security: almost half is vulnerable

The Veracode GenAI Code Security 2025 report analyzed code from over 100 large language models on 80+ real-world development tasks. The result: 45% of AI-generated code contains security vulnerabilities from the OWASP Top 10 list. Java turned out to be the worst language – over 70% failures. The most alarming conclusion of the report: larger and newer models did not solve the problem. The vulnerability level remained stable across model generations – GPT-4, GPT-5, Claude, Gemini. This is a structural problem inherent in how AI generates code, not a temporary limitation that will disappear with the next release.

"Our research shows that GenAI models make the wrong choice almost half the time — and it's not improving" — Jens Vessling, CTO of Veracode

Code quality: 1.7 times more defects

The CodeRabbit study "State of AI vs Human Code Generation" (December 2025) analyzed 470 real open-source pull requests on GitHub. AI-generated code contains 1.7 times more problems overall than human-written code – in every quality category: logic, maintainability, security, performance.

More details by category:

• Logical and correctness errors – 75% more often
• Security vulnerabilities – 1.5–2.74 times more often
• Code readability issues – more than 3 times more often
• Performance issues (excessive I/O, etc.) – almost 8 times more often

The average AI pull request contained 10.83 issues compared to 6.45 in human ones. Moreover, the issues in AI code are more serious: 1.4 times more critical and 1.7 times more major defects.

Real data leaks: 5000 applications without protection

In May 2026, the Israeli cybersecurity company RedAccess published the results of scanning 380,000 applications built on the Lovable, Replit, Base44, and Netlify platforms. Of these, over 5000 had practically zero protection and authentication. About 40% exposed sensitive data: medical records, financial documents, corporate materials, chat bot conversation logs. Confirmed leaks include the ship schedule of a British logistics company, clinical trial data of a British healthcare firm, and internal financial statements of a Brazilian bank.

The reason is structural: most of these platforms make new projects publicly accessible by default – and most non-technical users simply don't know that this needs to be changed. Security Boulevard wrote about this, and the original investigation was published in WIRED.

Business consequences: real losses for real companies

The Resume.org study (January 2026, 1146 American managers): 70% observed their subordinates making AI-related errors with real business consequences. And these are not isolated cases: 12% saw such errors "many times," 43% – "several times." Financial losses in individual cases exceeded $50,000. The most common errors were factual inaccuracies (58%) and ignoring critical context or nuances (50%).

These are not theoretical risks. These are real projects, real budgets, real consequences – documented in independent studies conducted in the first half of 2026.

Thousands of products. None memorable

There is a problem that is harder to measure than code vulnerabilities – but which is equally destructive to business. I noticed it not in research. I noticed it in my own work.

Lately, when I review new SaaS products, startup landing pages, or mobile apps – I increasingly get the feeling that I've seen it before. Not a specific product. A general feeling: the same page structure, the same "how it works" sections with three icons, the same hero text with "AI-powered" in the title, the same dashboard with a sidebar on the left and metric cards on top. As if someone created a template – and thousands of people replicated it.

This is not a coincidence. It is a direct consequence of how AI tools work.

AI models are trained on the same data – billions of lines of code, design systems, marketing texts that existed before their training. They gravitate towards the same libraries, the same UI patterns, the same phrasing. This is mathematically inevitable: the model optimizes for the "most probable" – and the most probable is the averaged.

The academic paper "Vibe Coding Kills Open Source" (arXiv, January 2026, authors Miklós Koren, Gábor Békés et al.) describes the mechanism precisely: in vibe coding, the AI agent itself chooses libraries and assembles code – often without the user reading the documentation or interacting with maintainers. Models gravitate towards large packages well-represented in the training data, which eliminates the organic process of tool selection. Researchers demonstrated this with Tailwind CSS: npm download counts increased – while Stack Overflow questions and documentation traffic decreased. Tailwind creator Adam Wathan wrote in a GitHub comment that his documentation traffic dropped by ~40% from its 2023 peak despite the framework's growing popularity – and revenue dropped by ~80%.

The same is happening with products: AI helps reach an average result faster – but it doesn't lead to an outstanding one. It doesn't know what makes your product memorable, because memorable is by definition not averaged.

I see this in WebsCraft's client projects too. When a client comes with a "reference" made through Lovable or v0 – it's functional, but completely soulless. There's no character, no voice, no sense that someone thought about the specific user. There's a feeling that someone formulated the prompt well.

The market feels this – even if it doesn't articulate it in words. A product that looks like all the others doesn't inspire confidence. Not because it's technically bad. But because it doesn't signal that there's a person behind it who thought, who has experience, who is responsible for the outcome.

In 2025, it was enough to build an application over the weekend through Claude or Cursor to generate interest – precisely because there were few such applications. In 2026, there are thousands of such applications. Speed to market has ceased to be a competitive advantage – it has become a minimum requirement. Differentiation is now entirely on the side of quality of thought, taste, and domain understanding. Things that AI doesn't generate – it only amplifies them in those who already have them.

AI is not to blame – the problem is with people without a foundation

It is important to stop here and say what many do not want to hear.

AI is not to blame.

Zehner and Ronacher – people who built a popular AI agent and actively use AI in development themselves – say directly: the problem is not with the tools. The problem is with how they are used. Zehner admits that AI tools are indeed useful for routine tasks. The issue is over-reliance on them – the belief that serious system design and testing can be replaced by a few well-formulated prompts.

Ronacher described two mechanisms he observes in serious projects:

• Automation bias – the tendency to trust the machine's output simply because it was generated by a machine, without critical analysis
• Review fatigue – when developers drown in AI-generated pull requests and gradually stop reading code carefully

Eric J. Ma – a developer who documented his own experience recovering from AI-generated chaos – put it precisely: "AI is an incredible executor. It can refactor, extract, implement. But vision? That remains human. And that vision comes from experience – from having seen enough codebases to know what works and what collapses under its own weight."

AI amplifies what is already in a person. If a developer or designer lacks taste, understanding of UX, architecture, or business – AI will simply help create a mediocre product faster. If they have it – AI becomes a true multiplier.

I myself described a similar situation – when my friend without a technical foundation decided to launch a product through Gemini over the weekend. In detail about where people without basic knowledge hit a wall – and why AI doesn't replace understanding architecture – I discussed in the article AI coding won't make you money. And here's why.

That is, the problem of vibe slop is not an AI problem. It is a problem of people without a foundation who have gained access to a powerful tool.

What's Next: A New Competitive Advantage

Zechner believes the reckoning is inevitable: large companies will soon realize that over-reliance on AI-generated code increases costs and degrades product quality. Many smaller startups dependent on vibe coding will fail. Cloud costs will rise because poorly written code consumes more computing resources, more memory, more traffic.

But there's a flip side. Those who learn to combine AI with fundamental skills correctly will gain an advantage that is difficult to copy.

The winners won't be those who generate code faster. The winners will be those who:

• Understand the user's problem — not superficially, but deeply: why they are looking for this solution, what truly bothers them, what will make them pay and return.
• Know how to design a product — architecture, UX, business logic: solutions that withstand load and don't crumble as the product grows.
• Have taste and expertise — the ability to distinguish "works" from "good"; understanding what makes a product memorable, not just functional.
• Control decisions, not delegate them to AI — AI executes, humans design and are responsible.
• Build brand and trust — in a world where thousands of identical products exist, recognition and reputation become a scarce resource.

As the author of the article on isquared.ie aptly put it: "AI doesn't remove the need to think. It removes the places where you can hide."

Previously, a mediocre product could be justified by the complexity of development. Now, that excuse is gone. AI has lowered the barrier to entry for everyone — which means differentiation now lies entirely with the quality of thinking, taste, and domain understanding.

The Main Takeaway

Vibe coding is not dead. Andrej Karpathy was not wrong — AI has indeed changed what it means to "know how to create software."

But the era of *mindless* vibe coding — when it was enough to quickly assemble a product and get it to market — is coming to an end. Not because AI has gotten worse. But because the market has adapted: what caused admiration a year ago is today the minimum bar.

The time has come for people who combine AI with professional skills. The time for those who have a foundation — understanding of architecture, UX, business logic, taste — and who use AI as a multiplier of that foundation, not as its replacement.

In 2025, the question was: "Can I do this with AI?"
In 2026, the question will be: "Do I have something that AI cannot provide — and am I using AI to amplify it?"