Carla Rover once spent thirty minutes sobbing after having to restart a project she had created using AI-generated code, known as vibe coding. With fifteen years of experience as a web developer, Rover is now building a startup with her son that develops custom machine learning models for marketplaces. She described vibe coding as a beautiful, endless cocktail napkin for sketching ideas. However, she noted that relying on AI-generated code for production can be worse than babysitting, as the models can introduce unpredictable errors.
Rover turned to AI coding tools for the promised speed in her startup development. In her rush to be quick and impressive, she took a shortcut by skipping a thorough scan of the files after an automated review. When she finally checked the code manually, she discovered numerous problems. A third-party tool revealed even more issues. This experience taught her a difficult lesson. She and her son ultimately had to restart their entire project, which led to her emotional reaction. She admitted she had handed off the work as if the AI copilot was a trusted employee, but learned that it is not.
Rover represents many experienced programmers who are turning to AI for coding assistance. However, these programmers often find themselves acting as AI babysitters, spending significant time rewriting and fact-checking the code that AI produces.
A recent report by content delivery platform company Fastly found that at least ninety-five percent of the nearly eight hundred developers surveyed spend extra time fixing AI-generated code. This verification burden falls most heavily on senior developers. These experienced coders have identified issues ranging from AI models hallucinating package names to deleting critical information and introducing security risks. Unchecked AI code can result in a product that is far buggier than one built by humans.
The problem has become so prevalent that it has given rise to a new corporate role known as the vibe code cleanup specialist. TechCrunch spoke with experienced coders about their use of AI-generated code and the future of vibe coding. While opinions varied, one point was certain: the technology still has a long way to go.
Rover compared using a coding co-pilot to giving a coffee pot to a smart six-year-old and asking them to pour coffee for the family. The child might succeed, but they could also fail, and most likely would not admit to any mistakes. She clarified that this does not make the child less clever, but means you cannot delegate such a task completely.
Feridoon Malekzadeh, who has over twenty years of experience in product development, software, and design, also compared vibe coding to working with a child. He is building his own startup and heavily uses the vibe-coding platform Lovable. For fun, he also uses AI to code apps, like one that generates Gen Alpha slang for Boomers. He appreciates the ability to work alone, saving time and money, but agrees that vibe coding is not like hiring an intern. Instead, he likens it to hiring your stubborn, insolent teenager to help. He said you have to ask them repeatedly to do something, and in the end, they do some of what you asked, some things you did not ask for, and break other things along the way.
Malekzadeh estimates he spends about fifty percent of his time writing requirements, ten to twenty percent on vibe coding, and thirty to forty percent on vibe fixing, which involves remedying bugs and unnecessary script created by the AI. He also noted that AI is not effective at systems thinking, the process of understanding how a complex problem impacts the overall result. AI-generated code tends to solve surface-level problems. For example, a good engineer would create a feature once for broad availability, whereas vibe coding might create the same feature five different times in five different places, leading to confusion for both the user and the model.
Rover finds that AI often runs into a wall when data conflicts with its hard-coded instructions. It can offer misleading advice, omit vital elements, or insert itself into a thought pathway you are developing. She also found that rather than admit errors, it will manufacture results. She shared an example where she questioned the results an AI model gave her. The model provided a detailed explanation pretending it had used the data she uploaded. Only when she called it out did the AI confess. She said it freaked her out because it sounded like a toxic coworker.
Security concerns are another major issue. Austin Spires, the senior director of developer enablement at Fastly, has been coding since the early 2000s. Through his own experience and conversations with customers, he found that vibe code prefers to build what is quick rather than what is right. This can introduce vulnerabilities similar to those made by very new programmers. He explained that an engineer often needs to review the code, correct the AI agent, and tell it that it made a mistake. This pattern is why the trope of AI models responding with “you’re absolutely right” when corrected has become common on social media.
Mike Arrowsmith, the chief technology officer at IT management software company NinjaOne, has been in software engineering and security for around twenty years. He stated that vibe coding is creating a new generation of IT and security blind spots, to which young startups are particularly susceptible. He explained that vibe coding often bypasses the rigorous review processes that are foundational to traditional coding and crucial for catching vulnerabilities. His company encourages safe vibe coding by using approved AI tools with access controls, mandatory peer review, and security scanning.
While nearly everyone agreed that AI-generated code and vibe-coding platforms are useful for mocking up ideas, they all stressed that human review is essential before building a business on it. Rover stated that the cocktail napkin is not a business model and that you must balance ease with insight.
Despite its errors, vibe coding has changed the present and future of the job. Rover said it helped her tremendously in crafting a better user interface. Malekzadeh said that even with the time spent fixing code, he still gets more done with AI coders than without them. He quoted French theorist Paul Virilio, who spoke about inventing the shipwreck along with the ship, noting that every technology carries its own negativity invented at the same time as technical progress. For him, the pros far outweigh the cons.
The Fastly survey found that senior developers were twice as likely to put AI-generated code into production compared to junior developers, saying the technology helped them work faster. Vibe coding is also part of Austin Spires’ routine. He uses AI coding agents on several platforms for personal projects, both front-end and back-end. He called it a mixed experience but said it is good for prototyping, building boilerplate, and scaffolding out tests. It removes menial tasks so engineers can focus on building, shipping, and scaling products. It seems the extra hours spent combing through AI-generated code will simply become a tolerated tax on using the innovation.
Elvis Kimara, a young engineer who recently graduated with a master’s in AI, is building an AI-powered marketplace. Like many coders, he said vibe coding has made his job harder and is often a joyless experience. He misses the dopamine rush from solving a problem himself, as the AI just figures it out. At his last job, he noticed senior developers were less inclined to help young coders; some did not understand the new vibe-coding models, while others delegated mentorship tasks to the AI models.
However, Kimara believes the pros far outweigh the cons, and he is prepared to pay the innovation tax. He said the new normal means developers will not just be writing code but will be guiding AI systems, taking accountability when things break, and acting more like consultants to machines. He plans to keep using AI even as he grows into a senior role, calling it a real accelerator. He makes sure to review every line of AI-generated code so he can learn even faster from it.