What happens when the next developer has to understand your AI code?

This study into the long-term cost of AI-driven development reveals some surprising findings. Most other studies only examine short-term outcomes. But what's the long-term impact of AI coding?

The real cost of AI code

Modern Software Engineering (MSE) recently conducted a study into the long-term effects of AI coding.¹

Their research was intentionally designed to explore the impact of AI coding in more realistic, real-world scenarios — such as what happens to your AI-driven code down the line.

There are lots of overly hyped AI claims. But are they looking at the right metrics? For instance:

1. AI finishes faster. Is that really the right thing to measure? Are we more concerned with “how fast a programmer types” versus the quality of their work?

2. AI code is more concise. This may sound good, but does “concise” also mean easy to understand, better documented or resulting in less maintenance?

Maintainability is important. It may be more important than the initial cost of development. Estimates vary, but it’s widely accepted that the actual cost of software is largely accrued from maintenance — about 3-4 times the cost of initial development.

If that’s correct, then I’d be more concerned about the effect on maintainability and less concerned about initial development.

Optimizing for short-term gains, like feature count, may be a silly trade-off.

The results are a bit surprising

As I mentioned, this was a serious study. MSE conducted what amounts to a blind study involving 151 participants. 95% of the participants were professional programmers in various industries — not just students.

The study focused on important topics and drew its findings from measurable metrics, including:

1. How long does it take to evolve AI-written code?

2. Are code quality metrics impacted?

3. Is test coverage affected?

4. Is productivity changed (as measured by the SPACE framework)?²

I won’t steal the big reveal from Dave Farley — it’s a short video (about 10 minutes) and loaded with good information. I recommend watching it.

I’ll offer a hint, though: developer skill might matter more than AI usage.

Fundamentally, this means AI code assistance can be an amplifier — meaning, it’s a double-edged sword. It amplifies both positive and negative practices. If you’re doing the right things, you get good results. But if you’re doing the wrong things, you’ll end up digging yourself into a hole even faster.

This emphasizes a message I’ve delivered before: AI tools can be a potential pitfall for junior developers. They need strong mentoring, enablement and guidance — not the kind that comes from an AI agent.

AI tools are potentially beneficial — but the benefit we get depends very much on the skill and expertise of the person using them.

If you’d like to read more, take a look at my article on planning your AI readiness.

Further reading

Planning ahead with an AI readiness plan: How AI will impact software development and security

1. Dave Farley, We Studied 150 Developers Using AI (Here’s What’s Actually Changed...), Jan. 27, 2026, Modern Software Engineering.

2. Microsoft, Developer Experience, Introducing the SPACE framework.