Generative AI tools are driving productivity across a wide variety of industries. The technology has found a lucrative niche in software development, where programmers report an up-to-30% improvement in output.

Vibe coding, as AI programming has come to be known, has the highest impact when used by more senior developers, allowing them to prototype and test ideas rapidly. But there are risks to this so-called “vibe coding” that development teams must be aware of, especially when code moves from experimentation and testing into production.

“AI-generated code is functional, but it often lacks optimisation, readability and long-term maintainability,” said Callan Abrahams, executive head of artificial intelligence at JSE-listed IT services group iOCO.

“It may solve an isolated issue but create hidden conflicts across system architectures, business logic or long-term strategy. Insecure patterns, flawed libraries and unvetted dependencies can easily slip into production. In other words, AI-generated code can solve one issue but introduce hidden conflicts elsewhere, because it can’t see the bigger picture and generated code may misalign with strict corporate architecture.”

According to Abrahams, vibe coding works well for proofs of concept because it is a powerful way to expand ideas and demonstrate potential to stakeholders quickly. However, migrating into a production environment requires strict controls that, if not present, increase the risk of “technical debt” in the system, especially for larger enterprises.

But there are risks outside of system integration requirements to consider, too.

While senior developers see AI tools as useful to cut out the mundane, routine and time-consuming aspects of their work, junior programmers who have not yet spent a significant portion of their careers wading through the weeds of software development are at risk of undertraining.

‘Understand the problem’

Russell Davidson, chief technology officer at BBD Software Development, told TechCentral that the number of iterations it takes to “vibe-code” a solution that fulfils requirements properly often takes longer than more conventional coding by skilled developers.

“We must ensure that junior engineers are still exposed to the fundamental building blocks of advanced coding techniques and patterns. A key principle is using AI to augment, not replace, existing skill sets. This means engineers should always understand the problem they are using the AI to solve and not use it as a substitute for a missing skill,” said Davidson.

Read: How AI is revolutionising computer programming

Vibe coding places a greater emphasis on the reading of code rather than the writing of it. According to Davidson, as important as reading code might be, when done together, reading and writing code places higher cognitive demands on the brain, helping improve retention and heighten skill.

iOCO’s Abrahams has a different view. She said the reading of code is a complex and layered activity.

“Historically, writing was the prized skill. Reading and debugging were assumed. With AI, reading, reasoning and architectural thinking become the premium skills. Developers must now validate AI outputs, spot inefficiencies and ensure maintainability. Skills in system-level understanding and teaching others to evaluate AI outputs are increasingly valuable in enterprise environments,” she said.

Abrahams said another risk posed by vibe coding is related to the tools being developed to enhance the developer experience, where programmers use existing AI tools to make even more specialised AI coding tools. “If an AI generates low-quality code for another AI system, the flaws are compounded and harder to detect. Debugging and governance become significantly more complex when AI-generated AI code isn’t properly understood,” she said.

The choice of tooling and how those tools are used is also important. Without a clear AI policy or well-defined AI stack, developers choosing their own AI models might opt for more generic utilities, whereas specialised models could produce better output.

Read: Wits professor named in Time’s list of 100 most influential people in AI

“We have found that mainstream, general-purpose models like ChatGPT, Claude and Gemini, and the agentic tools built on top of them like Cursor and Copilot, are generic and offer a ‘jack-of-all-trades’ level of output, whereas we can get better results with more finetuned or distilled models and agents built with LangGraph and LangChain,” said BBD’s Davidson. – © 2025 NewsCentral Media

Get breaking news from TechCentral on WhatsApp. Sign up here.