Carla Rover once found herself in tears for half an hour after being forced to redo a project she had created through vibe coding.
With 15 years of experience in the tech field, mostly as a web developer, Rover is now launching a startup with her son, focused on building bespoke machine learning solutions for online marketplaces.
She described vibe coding as an endless, creative space reminiscent of a cocktail napkin, always ready for new ideas. However, she pointed out that working with AI-generated code intended for production can sometimes be even more challenging than supervising a child, as these tools may introduce unpredictable errors.
Driven by the need to work quickly at her startup, she adopted AI coding to take advantage of the speed it promises.
“I wanted speed and to impress, so I cut corners and skipped reviewing the files after the automated checks,” she recalled. “When I eventually did a manual review, I discovered a lot of mistakes. Even more showed up with a third-party tool. That was a hard lesson.”
Ultimately, she and her son had to scrap their project and start over, which led to her emotional reaction. “I treated the copilot as if it were a staff member,” she admitted. “But it’s not.”
Rover is one of many veteran programmers who have turned to AI for assistance with coding. Yet these same developers often find themselves correcting, verifying, and rewriting the code produced by these AI tools.
A recent survey by Fastly, a content delivery network, showed that at least 95% of nearly 800 developers polled reported spending extra time fixing problems in code generated by AI, with senior engineers carrying most of this burden.
Seasoned developers have run into issues with AI code that range from invented package names to removal of vital information and security flaws. If unchecked, AI-generated code can leave products riddled with bugs beyond what a human might introduce.
The challenge of working with AI-written code has become so common that it has led to the emergence of a new corporate role: “vibe code cleanup specialist.”
TechCrunch interviewed experienced developers about their experiences with AI-generated code and their outlook on the future of vibe coding. While opinions differed, one consensus emerged: the technology still has significant room for improvement.
“Letting a coding co-pilot help is like handing a smart six-year-old a coffee pot and asking them to serve the family,” Rover explained.
Are they capable? Maybe. Could they make mistakes? Absolutely. And chances are, if something goes wrong, they won’t let you know. “It doesn’t mean the child isn’t smart,” she added. “It just means you can’t fully delegate such tasks.”
“You’re absolutely right!”
Feridoon Malekzadeh also likened vibe coding to working with a child.
With more than two decades in software, product, and design, Malekzadeh is now building his own startup and makes extensive use of the vibe-coding platform Lovable. He even enjoys creating apps for fun, like one that generates Gen Alpha slang for Boomers.
He appreciates the independence and cost savings vibe coding offers, but cautions that it’s not the same as hiring an intern or a junior developer. Instead, he says, it’s like “hiring your rebellious teenager to help out,” as he told TechCrunch.
“You have to repeat your instructions over and over,” he noted. “In the end, they complete part of the task, do some things you didn’t request, and break things along the way.”
Malekzadeh estimates about half his time is spent defining requirements, 10% to 20% on vibe coding itself, and 30% to 40% on “vibe fixing”—cleaning up bugs and removing unnecessary scripts produced by AI.
He also feels that vibe coding isn’t optimal for systems thinking—the ability to understand how complex issues affect overall outcomes. AI-generated code, he suggested, often focuses on superficial fixes.
“If you’re building a feature that needs to be accessible throughout your product, a skilled engineer would build it once and make it reusable,” Malekzadeh explained. “But with vibe coding, the same functionality might be recreated five separate times in five different ways if needed in different places, causing confusion for both users and the model.”
Meanwhile, Rover noticed that AI struggles when it comes across data that contradicts its pre-programmed logic. “It can give bad advice, omit crucial components, or derail your thinking process,” she said.
She has also observed that, rather than acknowledge mistakes, AI will sometimes fabricate results.
She recounted an incident to TechCrunch where she questioned the AI model’s output. The model responded with a detailed explanation, falsely claiming it utilized the data she provided. Only when she pressed further did it admit fault.
“It was unsettling—it reminded me of a toxic coworker,” she remarked.
Beyond these problems, security is also a concern.
Austin Spires, senior director of developer enablement at Fastly, has been programming since the early 2000s.
Through his own work and conversations with clients, he’s noticed that vibe coding often prioritizes speed over correctness, sometimes introducing vulnerabilities similar to those made by inexperienced coders.
“In many cases, the engineer must review the code, correct the AI, and point out its mistakes,” Spires told TechCrunch. “That’s why the phrase ‘you’re absolutely right’ has become a meme on social platforms.”
He’s referring to the tendency of AI models, such as Anthropic Claude, to reply with “you’re absolutely right” when their errors are pointed out.
Mike Arrowsmith, CTO of NinjaOne, an IT management software provider, has two decades of experience in software engineering and security. He noted that vibe coding is creating new IT and security blind spots, especially for early-stage startups.
“Vibe coding often skips the stringent review steps fundamental to traditional development and essential for catching vulnerabilities,” he told TechCrunch.
At NinjaOne, he said, they practice “safe vibe coding,” ensuring authorized AI tools have access controls and enforcing peer reviews and mandatory security scans.
The new normal
Most people interviewed agreed that while AI-generated code and vibe-coding platforms can be invaluable for prototyping or experimenting, human oversight remains vital before any business relies on them.
“That cocktail napkin isn’t a business plan,” Rover cautioned. “You need to pair convenience with critical thinking.”
Despite frequent complaints about its shortcomings, vibe coding has reshaped both present workflows and the future of the profession.
Rover shared that vibe coding greatly improved her approach to user interface design. Malekzadeh added that, even though he spends significant time correcting errors, AI coders still enhance his productivity compared to working alone.
Quoting French theorist Paul Virilio, Malekzadeh remarked, “‘Every technology comes with its own form of disaster, invented alongside the advancement itself.’”
The benefits far surpass the drawbacks.
Fastly’s study found that senior developers are twice as likely as juniors to deploy AI-generated code to production, citing increased speed as a key advantage.
Spires has also made vibe coding a regular part of his process, using AI agents on multiple platforms for both front- and back-end personal projects. He described the technology as a mixed bag, but praised its usefulness for prototyping, boilerplate setup, and scaffolding tests, freeing developers to focus on launching and scaling products.
It appears that investing extra hours in reviewing AI-generated code will simply become an accepted cost of leveraging the technology.
Elvis Kimara, a young engineer who recently earned a master’s in AI and is building an AI-driven marketplace, is learning this firsthand.
He admitted that, like many developers, vibe coding has made his role more difficult and sometimes less satisfying.
“The thrill of solving problems myself is gone—the AI does it for me,” he said. At a previous job, he noticed that senior developers provided less support to junior coders, either due to unfamiliarity with vibe-coding tools or because they left mentoring to the AI.
Still, he concluded, “the benefits far surpass the drawbacks,” and he’s ready to accept the price of innovation.
“Our roles are shifting—we’ll be supervising AI, taking responsibility for its errors, and acting as advisors to these systems,” Kimara said, describing the future he’s preparing for.
“Even as I advance to senior positions, I’ll continue using it. It’s been a genuine accelerator for my work, and by reviewing every line of AI code, I’m learning faster than ever.”
