The tech bubble is on fire right now. While some (Reddit cough, cough) rant about "AI slop", meaning low-quality AI code, others, myself included, are building apps at a speed that was unthinkable two years ago. I have been a developer for a long time and built software for years without AI. Today, though, there are five apps of mine in the App Store that wouldn't exist without AI support.
So where is this heading? Here is an attempt at an explanation in three acts, plus what follows from it in practice.
Where the term comes from, and what it doesn't mean
Andrej Karpathy coined the term in February 2025: vibe coding means fully giving in to the vibes and forgetting that the code even exists. You accept diffs without reading them, paste error messages back into the AI without comment, and find out whether it works by clicking the button. Karpathy explicitly meant it for throwaway weekend projects.
Simon Willison insisted early on that the term should not be watered down: if you read AI-generated code, understand it and can explain it to someone else, you are doing AI-assisted development. Vibe coding is the mode where exactly that doesn't happen. It sounds pedantic, but the distinction decides which of the following theses applies to a project.
The thesis: vibe coding frees the business departments
Until now, IT was the eternal bottleneck. Brilliant domain people such as logisticians, marketers or engineers had to painstakingly translate their ideas into requirement documents and hope the developers would understand them. Vibe coding breaks up this gatekeeping. The domain expert describes their intent, and the AI delivers a working prototype. That saves thousands of euros and months of time, and it's the ultimate democratization of software development: whoever understands the problem can now also build the solution. Or at least sketch out solution paths with precision.
Tools like Replit, Lovable and Bolt already sell exactly this promise: description in, app out. And a prototype built in an afternoon makes the better requirements document, because it shows what was meant instead of asserting it across forty pages.
The antithesis: we're breeding a technical time bomb
Just because a button clicks doesn't mean the app is good, though. When laypeople without any grasp of architecture "vibe" software into existence, they're building on sand. We risk a black hole of maintainability: code nobody understands anymore the moment the first bug shows up. Add to that massive security holes and systems that work with ten users but collapse at a hundred. Human laziness lures us into automation bias: we blindly trust the AI's output because it's convenient. And the AI produces plenty of nonsense and is often more show than substance.
I'm not going on gut feeling here; there are numbers. In 2025, Veracode had more than a hundred language models write code and tested the output against the OWASP Top 10: 45 percent of the code samples failed. In Java the failure rate was 72 percent, and in tasks where cross-site scripting was possible, the models skipped the defense in 86 percent of cases. The most uncomfortable finding: over time the models got better at writing working code, yet no better at writing secure code.
What that looks like in practice was documented in July 2025 by investor Jason Lemkin. During a multi-day vibe-coding experiment, Replit's agent deleted the production database with records on more than 1,200 executives and companies, even though a code freeze had been announced. It then denied it at first and generated thousands of fabricated entries that papered over the gap. Replit's CEO called the incident "unacceptable" and shipped safeguards afterwards. The actual lesson: the agent had access to production because nobody had separated it from production.
And automation bias is measurable. In a randomized study, METR had experienced open-source developers work on real issues from their own projects: with AI, they took 19 percent longer. They had expected to be 24 percent faster. Even afterwards, with their own stopwatch saying otherwise, they believed the AI had sped them up by 20 percent. The study is small and its setting specific, but the gap between felt and measured speed is exactly the raw material blind trust is made of.
The synthesis: from programmer to architect and auditor
The future belongs neither to the pure code typist nor to the clueless vibe coder. We're witnessing a radical role reversal: the senior developer becomes the adult in the room.
Much like an architect on a construction site, the developer of the future will do less bricklaying but will put their name on the line for the building's structural integrity and safety. We're moving away from "writing" toward "validating". A developer's value will be measured by judgment rather than by hours spent in front of an editor: knowing when an AI vibe is brilliant (or at least good enough) and when it poses a business-critical risk.
It's telling who trusts the AI least: in the 2025 Stack Overflow survey, 46 percent of respondents distrust the accuracy of AI tools while only 33 percent trust it, and the skepticism grows with professional experience. That is exactly the judgment the new role gets paid for. Whoever knows where AI code typically breaks reviews the right places.
What follows in practice
The failure patterns translate into rules that cost little:
- Treat AI code like someone else's code. A diff from the AI deserves the same review as an external pull request, especially when it runs on the first try.
- Wall off production. An agent doesn't need production credentials. The Replit incident wasn't a model failure; it was a missing boundary.
- Don't let the author grade the homework. Tests written by the same AI in the same context tend to confirm rather than check.
- Budget for security scans. The Veracode numbers don't improve because the prototype looks nice. A static scanner finds the usual suspects cheaply.
- Throw throwaway code away. The domain expert's prototype is an excellent requirements document. It is not a release candidate, and confusing the two is the most expensive mistake on this list.
Conclusion
Vibe coding is a powerful tool, but not a free pass. Karpathy described it for throwaway projects, and that's where it shines; everything beyond that is AI-assisted development and needs someone who understands the code. We need professionals more than ever. Their job is less typing and more curating a new, AI-assisted software world.
I originally shared this thought on LinkedIn (in German).
Sources
- Not all AI-assisted programming is vibe coding (but vibe coding rocks), Simon Willison, March 19, 2025
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, July 10, 2025
- We Asked 100+ AI Models to Write Code. Here's How Many Failed Security Tests, Veracode, 2025
- Stack Overflow Developer Survey 2025: AI, Stack Overflow, 2025
- Replit's AI Agent Wipes Company's Codebase During Vibecoding Session, Gizmodo, July 23, 2025