The Hidden Costs of Coding With Generative AI (MIT Sloan Management Review, August 18, 2025)

[highplainsdem]

I missed this when they published it 4 months ago, but they posted about it again on X two days ago:

https://sloanreview.mit.edu/article/the-hidden-costs-of-coding-with-generative-ai/

Generative AI can be a powerful productivity booster in coding — but only when deployed thoughtfully. Used carelessly, it can cripple scalability, destabilize systems, and leave companies worse off.

-snip-

When an organization rapidly introduces new software into existing systems, it can inadvertently create a tangle of dependencies that compounds its technical debt — that is, the cost of additional technological work that will be needed in the future to address shortcuts taken and quick fixes made during development. Technical debt is the hidden underbelly of digital technology. It is the 60-year-old COBOL code in banking systems that was never properly documented or updated. It is the shortcut of representing the current year with two digits instead of four, leading to the Y2K crisis, which cost hundreds of billions of dollars to fix globally. The buildup of technical debt causes slower development cycles, increased complexity, and security vulnerabilities, potentially leading to system failures.

-snip-

You can think of technical debt as operating much like financial debt. The “principal” is the work needed to modernize and refactor code; the “interest” is the ongoing complexity tax that slows maintenance, complicates scaling, and raises the risk of failure. While some debt is unavoidable, implementing AI-generated code is often akin to borrowing at a much higher interest rate. As one of the developers we interviewed said, “The problem with AI is that it can’t see the big picture.” Developers we interviewed also told us about code duplications, integration problems, dependency conflicts, a lack of context awareness, and myriad other problems that come with coding with AI. Indeed, when GitClear analyzed millions of lines of code from 2020 to 2024, it uncovered an eightfold increase in duplicated code blocks and a twofold increase in code churn — both measures of declining code quality. The “2024 Accelerate State of DevOps” report from Google’s DevOps Research and Assessment team found that a 25% increase in AI usage improves code review and documentation but results in a 7.2% decrease in delivery stability. So, what looks like rapid progress today could turn into costly setbacks tomorrow.

-snip-

Letting technical debt compound is dangerous. Southwest Airlines’ 2022 meltdown — which stranded over 16,900 flights and cost the airline over $750 million — was rooted in technical debt in its crew-scheduling systems. Technical debt drove the massive 2024 CrowdStrike outage that led to worldwide failures in health care delivery. In May 2025, Newark Liberty International Airport in New Jersey was plagued by massive delays and hundreds of flight cancellations that were caused by a combination of antiquated air traffic control technology and staffing shortages. Failures like these show how invisible risks can suddenly cripple even major organizations. Without deliberate efforts to “pay down the principal,” organizations risk becoming overwhelmed — first slowly, then all at once.

-snip-

I remember well-known software engineer Grady Booch posting on X about the technical debt problem with genAI coding over a year ago.

According to this article, the estimated cost of technical debt in the US is already over $2 trillion.

Trillion.

And AI coding is adding to that total cost rapidly.

The article warns that if inexperienced developers are deploying AI-generated code, and if that AI is being deployed in a brownfield (legacy) environment, "it may be best to avoid deploying AI-generated code entirely."

That's a pretty serious warning.

So we have the economy threatened by the AI bubble, and the environment and education threatened by AI as well.

And the coding that underpins much of the economy - and government, for that matter - is threatened by AI coding.

While the surface convenience of AI coding means fewer new developers will be educated, more and more developers will be laid off, and even employed developers will be deskilled by using AI.

But I suppose genAI peddlers and proponents will keep hoping that the hallucinating genAI endangering our coding now will somehow morph into the Magical Non-Hallucinating Coding Chatbot before this digital house of cards comes tumbling down.

[Diraven]

1. My wife teaches college computer science

So many of her students are using generative AI to do all their coding projects. And it's easy to catch because the code is full of unnecessary fluff that does nothing and very unoptimized code that barely completes its intended function. When she asks the students to explain how it works they have no idea. Also when they finally get to midterm and final exams they fail spectacularly because they learned almost nothing. She had to implement a policy that if students fail both the midterm and final then they automatically fail the class even if they get 100% on their homework. Last semester almost half the students failed her class like that.

[MineralMan]

2. AI also is not creative, but is reactive and prefers to use proven code.

That's a problem. At one time, I owned a small shareware software company. I had done all of the coding for the apps I was creating by myself. For various reasons, I was using Microsoft Visual Basic to create software for Windows. Some disputed that idea with me, but I found that it worked quite well for my purposes.

At one point, I decided that I might hire someone to do some of the coding, so I advertised for a Visual Basic programmer. Not many of those around, but I got a few applications. I narrowed it down to a couple of people and paid them at the advertised rate to test for the job.

The problem was a fairly basic one. I wanted to create a routine that would sort a medium--sized array that was in a table in the software. Basically, it was a simple address label database with five fields. I wanted to include a method for an alphabetical sort of the table based on any of the five fields, which were configured as an array, which I stored in memory in the application to increase the speed of operations. So, I assigned my applicants to write that sorting routing.

I had already done the job, and the sort was lightning fast. The table had no more than 750 rows and five columns.

After a longer time than I expected, the two applicants got it set up, each using a different array sorting process.

I tested their work with existing data. One was quite a bit faster than the other, but neither was anywhere near the speed of my routine. I hired nobody.

What they missed was that Visual Basic already had that sort built into the software. All you had to do was load the data into an actual table, which you could make invisible. Then, you could sort that table off-screen and replace the existing table with the sorted one. The whole process took less than 5 seconds. Microsoft's native sort was way faster than anything you could write in Visual Basic.

Neither of the applicants considered using an internal sorting tool, and wanted to show off their sorting routines. For me, the goal was "get it done and make it work fast."

I knew what Visual Basic had already done internally. The applicants didn't, so they wasted a huge amount of time writing their sorting routine, instead of using what was already coded by Microsoft.

Creativity. Doing things to meet goals rather than to be clever. I never did hire a coder, but continued to do it all myself.