AI Is Quick but Risky for Updating Old Software

Technical debt and hasty AI coding call for human oversight

Based on the research of Edward Anderson Jr.

At first glance, artificial intelligence looks like a software developer’s dream. A recent McKinsey & Company report found that programmers generated code up to 45% faster with the help of generative AI.

But if it’s not used strategically, AI can become a developer’s nightmare. So says Edward Anderson Jr., professor of information, risk, and operations management and Betty and Glenn Mortimer Centennial Professor in Business at Texas McCombs.

The problem arises when AI is used to write code that interacts with so-called legacy systems with outdated software, he explains. These environments are often awash in shortcuts, quick fixes, and other poor programming practices.

Such technical debt costs U.S. companies $1.5 trillion in reduced productivity and cybercrime, estimates the Consortium for Information & Software Quality.

It can also lead to real-world meltdowns. In 2022, Southwest Airlines’ 20-year-old scheduling system crashed, stranding passengers on nearly 17,000 flights.

But carelessly using AI to patch such systems risks making them even worse, Anderson warns. For one thing, AI trains on existing code, with all its defects. Thus, it tends to create more technical debt per line of code than trained, experienced human software engineers would.

How can companies avoid such problems? Anderson — with Geoffrey Parker of Dartmouth College and Burcu Tan of the University of New Mexico — interviewed dozens of programmers in a variety of industries. He offers some best practices for AI-assisted software development:

Make technical debt an engineering priority.

As companies race to market, fixing technical debt is often a low priority, Anderson says. They kick the can down the road, as Southwest Airlines did.

Rather than fixing something when it breaks, companies should incorporate the overhaul of technical debt into developers’ everyday workflows — especially if they’re using AI to make the repairs.

“This is about organizational processes,” he says. “If you’re going to use AI, and there’s a chance that you could be increasing the rate of technical debt generation, you’re going to have to allocate more time to doing the retirement.”

Develop clear guidelines for AI-assisted coding.

C-suite policies may address AI usage in general terms, but protocols in daily software development are still evolving, Anderson says. To help develop them, software teams should document when they are using AI and why.

Besides clearly defining tasks, they should also keep a human in the loop, he adds. “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.”

Train developers on the hazards of AI coding.

As the ranks of experienced developers shrink due to retirement, they may be replaced by inexperienced coders who deploy AI tools unchecked — especially in legacy environments.

In such situations, knowledge transfer is critical, Anderson says. Performance goals for senior programmers should include formal mentoring: not just reviewing code written by junior developers but training them in effective and responsible use of AI.

“Let me be clear: I think AI is a productivity booster,” Anderson says. “It’s just that you have to use it thoughtfully — and give software engineers the time to do that.”

“The Hidden Costs of Coding With Generative AI” is published in the MIT Sloan Management Review.

Story by Sally Parker