Coding Assistants Are Great—But Flawed

These days, venture capitalists, major artificial intelligence developers and cloud providers have zeroed in on AI-powered coding assistants as the killer product for large language models. 

But coding assistants aren’t perfect. It turns out that as they become ubiquitous, coding assistants are creating plenty of problems at companies that use them—mainly because junior programmers are relying too much on them.

There’s no question coding assistants are playing a bigger role in writing new code. Last week, for instance, Google CEO Sundar Pichai bragged about AI writing more than a quarter of all new code at his company. Microsoft has flaunted the fact that its GitHub Copilot coding assistant generates nearly half of the code in files in which software engineers enabled it. And OpenAI is doubling down on the code generation capabilities of its AI models, given Anthropic’s lead in the field and the fact that coding is one of the top reasons people subscribe to ChatGPT.

But let's not get too enthusiastic. Harish Mukhami, the CEO at GibsonAI, which sells AI software that can generate backend code, told me that coding assistants today are often missing a lot of context, such as information discussed during in-person meetings between engineers and product managers or an understanding of the broader project the assistant is creating code for.