Carla Rover once spent 30 minutes sobbing after having to restart a project she vibe coded. 

Rover has been in the industry for 15 years, mainly working as a web developer. She’s now building a startup, alongside her son, that creates custom machine learning models for marketplaces. 

She called vibe coding a beautiful, endless cocktail napkin on which one can perpetually sketch ideas. But dealing with AI-generated code that one hopes to use in production can be “worse than babysitting,” she said, as these AI models can mess up work in ways that are hard to predict. 

She had turned to AI coding in a need for speed with her startup, as is the promise of AI tools. 

“Because I needed to be quick and impressive, I took a shortcut and did not scan those files after the automated review,” she said. “When I did do it manually, I found so much wrong. When I used a third-party tool, I found more. And I learned my lesson.” 

She and her son wound up restarting their whole project — hence the tears. “I handed it off like the copilot was an employee,” she said. “It isn’t.” 

Rover is like many experienced programmers turning to AI for coding help. But such programmers are also finding themselves acting like AI babysitters — rewriting and fact-checking the code the AI spits out. 

A recent report by content delivery platform company Fastly found that at least 95% of the nearly 800 developers it surveyed said they spend extra time fixing AI-generated code, with the load of such verification falling most heavily on the shoulders of senior developers.

These experienced coders have discovered issues with AI-generated code ranging from hallucinating package names to deleting important information and security risks. Left unchecked, AI code can leave a product far more buggy than what humans would produce.

Working with AI-generated code has become such a problem that it’s given rise to a new corporate coding job known as “vibe code cleanup specialist.” 

TechCrunch spoke to experienced coders about their time using AI-generated code about what they see as the future of vibe coding. Thoughts varied, but one thing remained certain: The technology still has a long way to go. 

“Using a coding co-pilot is kind of like giving a coffee pot to a smart six-year-old and saying, ‘Please take this into the dining room and pour coffee for the family,’” Rover said. 

Can they do it? Possibly. Could they fail? Definitely. And most likely, if they do fail, they aren’t going to tell you. “It doesn’t make the kid less clever,” she continued. “It just means you can’t delegate [a task] like that completely.” 

“You’re absolutely right!” 

Feridoon Malekzadeh also compared vibe coding to a child.

He’s worked in the industry for more than 20 years, holding various roles in product development, software, and design. He’s building his own startup and heavily using vibe-coding platform Lovable, he said. For fun, he also vibe codes apps like one that generates Gen Alpha slang for Boomers. 

He likes that he’s able to work alone on projects, saving time and money, but agrees that vibe coding is not like hiring an intern or a junior coder. Instead, vibe coding is akin to “hiring your stubborn, insolent teenager to help you do something,” he told TechCrunch. 

“You have to ask them 15 times to do something,” he said. “In the end, they do some of what you asked, some stuff you didn’t ask for, and they break a bunch of things along the way.” 

Malekzadeh estimates he spends around 50% of his time writing requirements, 10% to 20% of his time on vibe coding, and 30% to 40% of his time on vibe fixing — remedying the bugs and “unnecessary script” created by AI-written code. 

He also doesn’t think vibe coding is the best at systems thinking — the process of seeing how a complex problem could impact an overall result. AI-generated code, he said, tries to solve more surface-level problems. 

“If you’re creating a feature that should be broadly available in your product, a good engineer would create that once and make it available everywhere that it’s needed,” Malekzadeh said. “Vibe coding will create something five different times, five different ways, if it’s needed in five different places. It leads to a lot of confusion, not only for the user, but for the model.”

Meanwhile, Rover finds that AI “runs into a wall” when data conflicts with what it was hard-coded to do. “It can offer misleading advice, leave out key elements that are vital, or insert itself into a thought pathway you’re developing,” she said. 

She also found that rather than admit to making errors, it will manufacture results.

She shared another example with TechCrunch, where she questioned the results an AI model initially gave her. The model started to give a detailed explanation pretending it used the data she uploaded. Only when she called it out did the AI model confess.

“It freaked me out because it sounded like a toxic co-worker,” she said.

On top of this, there are the security concerns.

Austin Spires is the senior director of developer enablement at Fastly and has been coding since the early 2000s. 

He’s found through his own experience — along with chatting with customers — that vibe code likes to build what is quick rather than what is “right.” This may introduce vulnerabilities to the code of the kind that very new programmers tend to make, he said. 

“What often happens is the engineer needs to review the code, correct the agent, and tell the agent that they made a mistake,” Spires told TechCrunch. “This pattern is why we’ve seen the trope of ‘you’re absolutely right’ appear over social media.” 

He’s referring to how AI models, like Anthropic Claude, tend to respond “you’re absolutely right” when called out on their mistakes.

Mike Arrowsmith, the chief technology officer at the IT management software company NinjaOne, has been in software engineering and security for around 20 years. He said that vibe coding is creating a new generation of IT and security blind spots to which young startups in particular are susceptible.

“Vibe coding often bypasses the rigorous review processes that are foundational to traditional coding and crucial to catching vulnerabilities,” he told TechCrunch.

NinjaOne, he said, counters this by encouraging “safe vibe coding,” where approved AI tools have access controls, along with mandatory peer review and, of course, security scanning. 

The new normal

While nearly everyone we spoke to agrees that AI-generated code and vibe-coding platforms are useful in many situations — like mocking up ideas — they all agree that human review is essential before building a business on it.

“That cocktail napkin is not a business model,” Rover said. “You have to balance the ease with insight.” 

But for all the lamenting on its errors, vibe coding has changed the present and the future of the job. 

Rover said vibe coding helped her tremendously in crafting a better user interface. Malekzadeh simply said that, despite the time he spends fixing code, he still gets more done with AI coders than without them.

“‘Every technology carries its own negativity, which is invented at the same time as technical progress,” Malekzadeh said, quoting the French theorist Paul Virilio, who spoke about inventing the shipwreck along with the ship.

The Fastly survey found that senior developers were twice as likely to put AI-generated code into production compared to junior developers, saying that the technology helped them work faster. 

Vibe coding is also part of Spires’ coding routine. He uses AI coding agents on several platforms for both front-end and back-end personal projects. He called the technology a mixed experience but said it’s good in helping with prototyping, building out boilerplate, or scaffolding out a test; it removes menial tasks so that engineers can focus on building, shipping, and scaling products. 

It seems the extra hours spent combing through the vibe weeds will simply become a tolerated tax on using the innovation.

Elvis Kimara, a young engineer, is learning that now. He just graduated with a master’s in AI and is building an AI-powered marketplace. 

Like many coders, he said vibe coding has made his job harder and has often found vibe coding a joyless experience. 

“There’s no more dopamine from solving a problem by myself. The AI just figures it out,” he said. At one of his last jobs, he said senior developers didn’t look to help young coders as much — some not understanding new vibe-coding models, while others delegated mentorship tasks to said AI models.

But, he said, “the pros far outweigh the cons,” and he’s prepared to pay the innovation tax. 

“We won’t just be writing code; we’ll be guiding AI systems, taking accountability when things break, and acting more like consultants to machines,” Kimara said of the new normal for which he’s preparing.  

“Even as I grow into a senior role, I’ll keep using it,” he continued. “It’s been a real accelerator for me. I make sure I review every line of AI-generated code so I learn even faster from it.”