AI Writes 30% of Our Code. So What?

There’s a new favorite line among Big Tech CEOs: “AI writes 30% of our code.” It sounds impressive. But sniff around a little, and you catch a familiar scent. It’s lines of code — the metric software engineering buried decades ago — back from the dead, dressed up in an AI costume.

Let me explain why this bragging contest is a warning sign, and what we should actually be measuring instead.

What “AI Writes 30% of Our Code” Really Means

Over the past couple of years, a parade of tech leaders has made nearly identical claims. AI generates a quarter of their internal code. At some companies, nearly half. Read it cold and it sounds like a revolution. But one question deserves an answer first: how exactly is that percentage measured?

In most cases, it’s the number of characters autocomplete suggested, or the lines an AI tool nudged into a commit. In other words, the volume AI typed. Whether the code was any good, whether it actually shipped, whether it survived a month in production — none of that gets counted.

Here’s the heart of it. This metric measures output, not productivity. Those are two completely different stories.

The Return of LOC, the Metric Everyone Mocked

Lines of code — LOC — is a long-running embarrassment in software. Since the 1980s, managers have tried to grade engineers by how many lines they wrote, and it has failed spectacularly every single time.

The reason is simple. Measure people by line count, and people optimize for line count. They stretch a three-line solution into ten. They copy-paste duplicate logic instead of consolidating it. Good engineers do the opposite — they shrink code, merge it, delete it.

Which brings up the famous joke: measuring programming progress by lines of code is like measuring aircraft building progress by weight. The line, often attributed to Bill Gates, nails exactly how hollow LOC is.

Then the AI era arrived, and the ghost quietly came back to life. This time it’s not grading individual engineers. It’s a trophy for the whole company’s “AI adoption.”

Why CEOs Cling to This Number

To get it, stand in their shoes. Executives who poured enormous sums into AI need proof the bet is paying off. And the easiest number to wave at shareholders and the market is exactly this: “AI generates X% of our code.”

The trouble is that the number is perfect for show and empty on substance.

First, the yardsticks don’t match. Nobody knows whether Company A’s 30% means the same thing as Company B’s 30%.

Second, it’s unclear the AI code is even used. An engineer accepts an autocomplete suggestion, deletes it ten seconds later — and it still shows up in the count.

Third, and most important, it misses the point entirely. Less code is frequently better code. A clean 100 lines beats a buggy 1,000 every time.

The Metrics That Actually Matter

So how do you know whether AI is genuinely helping? You look at outcomes, not line counts.

Are features shipping faster? Are there fewer bugs? Are engineers freed from tedious busywork so they can focus on the hard, important problems? That’s real productivity. It’s harder to measure — but “hard to measure” is no excuse for measuring the wrong thing instead.

And on the ground, plenty of engineers are uneasy. Reviewing and fixing the flood of AI-generated code takes serious time. If generation got faster but review got heavier, overall velocity can actually drop. That’s code debt, quietly compounding.

The Takeaway

“Percentage of code written by AI” is a great marketing line. But underneath, it’s a reheated version of a metric that failed 40 years ago. We already learned — at considerable cost — that a bigger number doesn’t mean better work.

So the next time a company brags that AI writes half its code, try asking one question back: did the software actually get better? Bragging about a number and creating value are not the same thing. Which one are you watching?