Carla Rover once spent 30 minutes sobbing after he had to restart her coding project.
Rover has been in the industry for 15 years and is primarily a web developer. She is now building a startup with her son that creates customized machine learning models for the market.
She calls the atmosphere coded as beautiful, endless cocktail napkins where people can draw ideas forever. But dealing with AI-generated code that one wants to use in production is “worse than a nanny”, she said, because these AI models can mess up work in unpredictable ways.
She has turned to AI coding so that, like the promise of AI tools, needs to be used with her startup at a speed.
“Because I needed to be fast and impressive, I chose the shortcuts and didn’t scan the files after automated review,” she said. “When I did this manually, I found a lot of errors. When I used third-party tools, I found more. I learned lessons.”
She and her son ended the entire project reboot – so shed tears. “I handed it to it like a co-pilot was an employee,” she said. “No.”
Rover, like many experienced programmers, turns to AI for coding help. However, such programmers also find themselves behaving like AI nanny – rewriting and fact checking the code that AI spits out.
TechCrunch Events
San Francisco
|
October 27-29, 2025
The latest report from content delivery platform companies is fast At least 95% were found Of the nearly 800 developers surveyed, they spent the extra time repairing AI-generated code, the biggest load on this verification fell on the shoulders of senior developers.
These experienced coders have identified problems with AI-generated code, from hallucinating packaging names to removing important information and security risks. Unchecked AI codes may make products larger than the larger goods produced by humans.
Code generated using AI has become so problematic that it has caused new corporate coding efforts, called “Ambient code cleaning expert. ”
TechCrunch uses AI-generated code to talk to experienced coders about their time, involving what they think is the future of Vibe encoding. The idea is different, but one thing is still certain: this technology still has a long way to go.
“Using a co-pilot with a co-pilot is like building a coffee pot for a smart six-year-old and saying, ‘Please bring it into the restaurant and pour coffee for the family.”
Can they do it? possible. Will they fail? really. And most likely, if they do fail, they won’t tell you. “It won’t make the kids less smart,” she continued. “It just means you can’t delegate [a task] That’s exactly. ”
“You are absolutely right!”
Feridoon Malekzadeh also compared the coding atmosphere with the children.
He has worked in the industry for over 20 years, holding various roles in product development, software and design. He said he is building his own startup and using the Vibe encoding platform cutely. For entertainment, he also used vibe code that generates Gen Alpha language for Boomers.
He likes being able to work on projects alone, saving time and money, but agrees that atmosphere coding is not like hiring an intern or junior coder. Instead, he told TechCrunch that Vibe codes are similar to “hiring your stubborn, rude teenager to help you do something.”
“You have to ask them to do something 15 times,” he said. “In the end, they did some of the things you asked for, some of the things you didn’t ask for, and broke a lot of things along the way.”
Malekzadeh estimates that his time accounts for about 50% of the time, with 10% to 20% of the time, 30% on Vibe encoding, and 30% to 40% of his time on Vibe fixing – Remediate errors and “unnecessary scripts” created by AI encoding.
He also believes that Vibe encoding is the best encoding for system thinking, that is, the process of looking at how complex problems affect the overall results. He said AI-generated code tries to solve more surface-level problems.
“If the product you are creating should be widely used in the product, a good engineer will create it once and use it anywhere,” said Malekzadeh. “If you need it in five different places, Vibe encoding will create five different times, five different ways. It will cause a lot of confusion, not only for the user, but for the model.”
Meanwhile, Rover found that when the data conflicted with hard-coded data, the AI ”hit to the wall”. “It can provide misleading advice, ignore crucial key elements, or insert yourself into the path of thought you are developing,” she said.
She also found that it does not admit to making mistakes, but produces results.
She shared another example with TechCrunch, where she questioned the results the AI model initially gave her. The model begins to provide a detailed description of pretending to use the data she uploads. Only when she summons it, the AI model admits it.
“It sounds like a toxic colleague, and it shocks me,” she said.
Most importantly, there are security issues.
Austin Spiers is the fast-growing senior director of developer support and has been coding since the early 2000s.
Through his own experience (and chatting with clients) he discovered that Vibe Code likes to build fast rather than “right”. This could introduce vulnerability into very new codes that programmers tend to make, he said.
“What happens often is that engineers need to look at the code, correct the agent, and tell the agent that they made a mistake,” Spiers told TechCrunch. “This pattern is why we see the ‘you’re absolutely right’ telescope that appears on social media.”
He refers to AI models, such as the anthropomorphic Claude, which tends to respond “absolutely correct” when they make mistakes.
Mike Arrowsmith, chief technology officer at Ninjaone, IT management software company, has been engaged in software engineering and security for 20 years. He said Vibe coding is creating a new generation of IT and security blind spots, especially when young startups are vulnerable.
“Vibe encoding often bypasses rigorous scrutiny processes that are the basis of traditional encoding and are crucial to catching vulnerabilities,” he told TechCrunch.
He said the Ninja countered the issue by encouraging “safe Vibe encoding”, where approved AI tools have access controls, as well as mandatory peer review and of course secure scanning.
The new normal
Although almost everyone we talked to agree that AI-generated code and Vibe encoding platforms are useful in many cases, such as ridiculing ideas, they all agree that human comment is crucial before building a business.
“That meal napkin is not a business model,” Rover said. “You have to strike a balance between insight and insight.”
But for all the lament of errors, Vibe encoding changes the present and future of work.
Rover said Vibe coding has greatly helped her make a better user interface. Malekzadeh just said that despite his time fixing the code, he was more done with the AI coder than without them.
“Each technology has its own negative emotions, which was invented at the same time as technological advances.” Malicard quoted French theorist Paul Virilio, who spoke about the idea of inventing a shipwreck with the ship.
The advantages far outweigh the disadvantages.
The quick survey found that AI-generated code is twice as likely to be put into production than junior developers, saying the technology helps them work faster.
Vibe encoding is also part of the Spiers encoding program. He uses AI encoding agents on several platforms for front-end and back-end personal projects. He calls the technology a hybrid experience, but says it is great in helping with prototypes, making templates or scaffolding testing. It removes trivial tasks so engineers can focus on building, transporting and expanding products.
It seems that the extra time spent on Vibe weeds will only become tolerant taxes using innovation.
Young engineer Elvis Kimara is now learning. He has just graduated from a master of AI and is building an AI-driven market.
Like many coders, Vibe coding makes his work harder and often finds ambience coding a joyless experience, he says.
“I solved the problem myself. AI just figured it out,” he said. In his last job, senior developers didn’t hope to help young coders too many – some don’t understand the new atmosphere coding model, while others don’t understand the guiding tasks of AI models.
But, he said: “Professionals far outweigh the shortcomings” and he is ready to pay the innovation tax.
“We’re not only going to write code; we’re going to guide the AI system, take responsibility when things break down, and be more like a consultant to machines,” Kimara said of the new normal he prepared for.
“I’ll keep using it even if I’m a senior position,” he continued. “It’s a real accelerator for me. I make sure to look at every line of code generated by the AI, so I learn from it even faster.”
