Day: August 28, 2025

• Vibe coding is sweeping through Silicon Valley with more and more AI tools that help with coding.
• Computer scientists say coding with AI is fun and fast, but that it won’t replace engineers.
• What it means to be a software enginjeer is evolving with advancements in generative AI.

Kevin Wu, CEO and cofounder of the AI voice agent startup Leaping AI, said everyone is vibe coding.

But that doesn’t mean they’re all doing it well.

Wu said his company hired an engineer, only to suspect they were mostly vibe coding. “They didn’t even have any idea what the code does that the AI produced,” he said.

Vibe coding can be defined in different ways, but it generally refers to using plain English to instruct AI tools to write code. The phenomenon has swept through the tech world, calling into question the wisdom of the widespread career advice that simply said: “learn to code.”

Still, engineers at major tech companies and computer science professors told Business Insider that while AI can make coding faster and more fun, it’s no replacement for software engineers.

“It’s sort of like giving a novice driver a race car,” Nenad Medvidovic, chair of the Department of Computer Science at the University of Southern California, said. “There are certain things that can go very wrong if you don’t know what you’re doing.”

Medvidovic said studies show that AI works well for easier or moderately complex pieces of code, but that once you give it more complex problems, “vibe coding becomes less useful.”

Coding with AI has its limits

The rise of AI could already be affecting entry-level coding jobs. Researchers at Stanford found in a new study that in the most AI-exposed professions, which they said included software engineers, workers ages 22 to 25 saw a 13% relative decline in employment since the widespread adoption of generative AI. Employment for early career software developers declined by nearly 20% from its peak in late 2022 to July 2025, the study found.

Todd Millstein, a computer scientist and professor at the University of California, Los Angeles, said coding with AI is similar to coding with a junior engineer. You give it instructions and say, “Why don’t you try this?” and then you might review it and say, “This part is good, but try doing this part differently.”

However, the industry is a long way from the common misconceptions of vibe coding, where the person doing it doesn’t need to be able to understand code, and the AI can handle a project from beginning to end.

“You’re reading the code. You’re responsible for the code,” Millstein said of engineers using AI tools. “And sometimes there are parts that are just too tricky or would be too complicated to explain to the AI, and so you end up doing it yourself as well.”

Millstein said the initial code creation is also a small part of the software cycle — it also needs to be maintained and has to have bug fixes and extensions over time.

“It definitely is not the same as what a software engineer is doing or needs to do,” he said.

He also noted that when it comes to the systems our world depends on — financial institutions, the stock market, healthcare, medicine — the stakes are simply too high to trust AI-generated code without a person who actually understands the code, security requirements, and how it interacts with other parts of the system.

Jigar Bhati, an engineer at OpenAI, said AI tools act more as an assistant rather than automating the process of coding. As the human engineer, he said, “You’re still in control.” He also said that vibe coding can be useful for creating a quick prototype or proof of concept, but to be a usable product, it requires experienced engineers with a critical understanding of how the system works.

“And there’s still an aspect of, you need to talk to 10 different teams to make a great product, and that is something software engineers are involved in day-to-day life as well, apart from just coding,” he said.

Ritvika Nagula, a software engineer at Microsoft, said having a deeper understanding of code is necessary to know how to prompt the AI tools well. AI responses are often only as good as the prompts they are given. For instance, she said if you don’t give the AI tool the right context, “it might give you a solution that’s not entirely applicable.”

Vibe coding is changing the nature of software engineering

Despite its limitations, the engineers and professors said vibe coding can make programming faster and more fun.

“You can experiment with things very quickly,” Antara Dave, a product designer at Microsoft, said. “It’s very fun as well. You can generate a lot of good ideas within seconds or minutes. You can go from one solution to the other solution.”

Millstein said AI is currently pretty good at doing some of the basic coding that every app needs, which speeds up the process and allows engineers to focus on the actually fun and challenging parts. “Just taking away the mundane stuff is a huge productivity gain,” he said.

AI tools are improving rapidly, and Millstein and Medvidovic said they are constantly re-evaluating what that means for software engineers and computer science curricula.

“One of the things that we recognize is that our students are going to use it whether we want them to or not,” Medvidovic said of AI, adding that in addition to teaching the fundamentals of computer science, they also need to teach students how to use AI tools responsibly.

The professors said they are still focusing on instilling fundamentals, which they don’t see as becoming unnecessary anytime soon.

Rather than replace engineers, they said it’s more likely that what it means to be a software engineer will continue to evolve, and will no doubt include vibe coding AI, whatever form it ultimately takes.