The Reality of Vibe Coding

In a late-night project overhaul, Carla Rover, a developer with 15 years of experience, cried for half an hour. It wasn’t due to a bug but because she thought she had found a secret weapon in AI-assisted coding, only to end up spending more time cleaning up the mess.

This experience is not unique.

With the rise of AI tools like GitHub Copilot, ChatGPT, and Cursor, many seasoned programmers have joined the trend of “Vibe Coding”—throwing ideas at AI to generate code while checking, fixing, and even rewriting it line by line. While AI appears to be a helpful assistant, many have found themselves becoming “AI babysitters.”

From Clean Slate to Restart

Rover, who primarily worked in web development, is now trying to build a customized machine learning model for an e-commerce platform with her son. She described Vibe Coding as “a clean sheet of paper to doodle ideas on.”

However, once she put AI-generated code into production, problems arose.

To meet deadlines, she initially relied completely on AI for automated reviews, skipping manual checks. When she later reviewed the code, she was shocked by the number of bugs, and third-party tools also reported errors. Ultimately, she and her son had to restart the entire project.

“I thought of Copilot as a reliable employee, but it proved otherwise.”

This experience aligns with a recent Fastly survey, which found that 95% of nearly 800 developers reported needing extra time to modify AI-generated code, with most of the fixing work falling on senior engineers.

The issues with AI-generated code are varied: fictitious package names, deletion of critical information, and security vulnerabilities. If unchecked, AI-generated code can be more fragile and bug-ridden than hand-written code. These serious problems have even led to the emergence of a new role—“Vibe Code Cleanup Specialist.”

A Day in the Life of an AI Babysitter

Similarly, another seasoned developer, Feridoon Malekzadeh, has had a complex experience.

With over 20 years in product development, software, and design, he also uses Vibe Coding platforms like Lovable extensively. He has even created some “toy” applications, such as one that generates slang for the Baby Boomer generation.

While it sounds fun, the reality feels like hiring a rebellious teenager:

“You have to prompt it multiple times before it reluctantly does something. In the end, it only completes part of the request and adds a bunch of unwanted things that break other features.”

He estimates his time allocation roughly as follows:

• 50% spent writing requirements
• 10-20% letting AI write code
• 30-40% fixing bugs and redundant code generated by AI

In other words, the time saved through Vibe Coding is minimal.

Even more frustrating is that AI lacks systematic thinking. An experienced developer might write a general module for reuse, while AI might create five different implementations in five places, increasing maintenance costs and complicating the project.

AI’s Denial and Security Risks

In addition to frequent bugs, Rover noticed another unsettling phenomenon: when AI encounters data conflicts, it not only fails to acknowledge errors but also fabricates results.

For instance, when she questioned the logic of a piece of AI-generated code, the model began to “explain” that it used the uploaded data. Only when confronted did it admit, “Actually, it didn’t.”

“At that moment, I felt I was dealing with a ’toxic colleague’ rather than a tool.”

In fact, AI security risks are a concern in the industry. Austin Spires, Fastly’s Director of Developer Empowerment, noted that AI often prioritizes “speed” over “accuracy,” frequently introducing bugs that only beginners would make.

This is why social media often features the joke that “AI always replies ‘You’re absolutely right’"—developers point out errors, and AI immediately “admits fault,” but the previous responses were already incorrect. Mike Arrowsmith, CTO of NinjaOne, warns that using Vibe Coding can easily bypass traditional code review and security processes, especially in startups.

Despite the myriad problems with Vibe Coding, nearly all developers admit that AI coding is still indispensable. It is particularly suited for prototyping, quick mocks, generating templates, or testing tasks, significantly reducing repetitive labor.

As French theorist Paul Virilio said, “While building ships, we also invented shipwrecks.” In Malekzadeh’s view, the various downsides of AI coding are also a byproduct of progress.

Moreover, Fastly’s survey results show that senior developers are twice as likely to put AI code into production compared to junior developers—indicating that while they spend considerable time modifying AI code, their experience allows them to utilize this technology effectively.

The Lost Joy of Programming for the Younger Generation

Unlike seasoned developers who invest in and affirm AI coding, younger engineers feel they have lost much of the joy of programming.

For example, Elvis Kimara, a recent AI master’s graduate developing an AI-driven e-commerce platform, admits that Vibe Coding has diminished his sense of accomplishment: “The dopamine from solving problems myself is gone; AI just takes care of it.”

He also observed that some senior developers, after using AI, have reduced their help to newcomers. Some even shift the responsibility of mentoring to AI, while others do not fully understand how the new tools operate.

However, Kimara does not reject AI: “The benefits outweigh the drawbacks, and I’m willing to pay the price for this innovation. Future developers will not just write code but will guide AI and take responsibility for errors, resembling an AI consultant role.” He emphasizes that even as a senior developer, he will continue to use AI while meticulously reviewing AI-generated code to learn more.

Undoubtedly, Vibe Coding is quietly changing the way developers work.

It is not a perfect tool, nor is it a “zero-cost productivity multiplier”; instead, the bugs, redundancies, risks, and responsibilities it brings are becoming a form of “innovation tax” that developers must bear. Yet at the same time, it accelerates project delivery and expands the boundaries for individual developers and small teams.

Thus, for many developers, being an “AI babysitter” is hard work but worth it. What are your thoughts?