Artificial intelligence is helping programmers write code faster than ever before, but a new study warns that this speed comes at a dangerous price when dealing with the outdated software that underpins much of the corporate world.
Research from the University of Texas at Austin suggests that while generative AI is a “productivity booster”, its careless application to legacy systems could trigger disastrous meltdowns.
“At first glance, artificial intelligence looks like a software developer’s dream,” said Edward Anderson Jr., a professor at the McCombs School of Business. He notes that recent reports show AI can speed up coding by up to 45 per cent.
“But if it’s not used strategically, AI can become a developer’s nightmare.”
The ‘technical debt’ trap
The core of the problem lies in “technical debt”, the accumulation of shortcuts, quick fixes, and poor programming practices that inevitably accumulate in legacy software systems over decades. This debt already costs US companies an estimated $1.5 trillion (£1.1 trillion) annually in lost productivity and cybercrime damages.
Anderson warns that using AI to patch these rickety systems often makes them worse. Because AI models are trained on existing code — including flawed code with defects and shortcuts — they tend to replicate these errors.
“Thus, it tends to create more technical debt per line of code than trained, experienced human software engineers would,” Anderson explained.
The consequences of ignoring this debt can be severe. The study cites the 2022 collapse of Southwest Airlines’ 20-year-old scheduling system, which left passengers stranded across nearly 17,000 cancelled flights, as a cautionary tale of what happens when legacy infrastructure fails.
Humans in the loop
To prevent similar catastrophes exacerbated by AI, Anderson and his colleagues—Geoffrey Parker of Dartmouth College and Burcu Tan of the University of New Mexico—interviewed dozens of programmers to develop a set of safety protocols.
They argue that companies must stop “kicking the can down the road” and make the retirement of technical debt a daily engineering priority, rather than fixing things only when they break.
Crucially, the study calls for strict human oversight. As experienced developers retire, there is a risk that junior coders will deploy AI tools without proper safeguards. The researchers advise that senior staff must formally mentor junior developers on the specific hazards associated with AI coding.
“You really want to make sure you’ve got somebody who has a lot of training in software engineering and experience to catch the AI when it’s making mistakes,” Anderson said.
“Let me be clear: I think AI is a productivity booster. It’s just that you have to use it thoughtfully — and give software engineers the time to do that.”
