Google's New Metric: 75% of Code Written by AI. Should Programmers Worry?

Last Wednesday, Google’s CEO Sundar Pichai published a blog post saying that 75% of the company’s newly written code is now AI-generated. My first reaction: wow.

Honestly, this number came faster than I expected. I always thought “AI writing code” was mostly media hype, but Google isn’t some small startup — it’s a company with tens of thousands of engineers, and they’re genuinely handing 75% of new code to AI. That’s a different story entirely.

Here’s the context. Google started pushing AI coding tools back in 2024, initially as an assistant — AI writes a version, humans review it. Less than two years later, the ratio went from 25% to 50% to now 75%. At this pace, it won’t be long before the model flips entirely: humans give instructions, AI does everything else.

A friend of mine works backend at ByteDance. He told me their team’s process for a new feature module is now: PM raises a request → programmer describes the logic clearly → AI generates the code → programmer reviews it → ships. His exact words: “Writing code now feels more like writing prompts than writing code.”

That resonated with me. I’ve been doing some AI consulting lately, and I sometimes find myself wondering: where exactly is my value in all this?

I think it’s important to pour some cold water here. The 75% figure sounds like programmers are done for, but actually, this number reveals something else: the value of human engineers has shifted from “writing code” to “defining problems,” “reviewing logic,” and “architecting systems.” AI can write code, but it doesn’t know why this particular piece of code should exist, who it’s serving, or whether the requirements themselves are flawed.

In other words: AI ate the “execution layer” of programming, but the “decision layer” and “judgment layer” still belong to humans. This is similar to self-driving cars — Level 2 autonomous driving gets hyped to the heavens, but true fully driverless cars are still a long way off.

But here’s something worth noting: this “AI writes code + human reviews” model actually raises the bar for programmers, not lowers it. Previously, you just needed to “know how to code.” Now you need to “ask the right questions,” “spot AI’s mistakes,” and “judge from a business perspective whether the output makes sense.”

For newcomers with less than three years in the industry, this is actually tough. You don’t have enough experience to judge whether AI-generated code is correct or optimal. You might fall into the trap of thinking “if AI wrote it, it must be right.”

So my advice: if you’re still learning to code, don’t bet everything on “learning to write code.” What matters more is learning to “understand problems,” “break down requirements,” and “validate results” — skills AI can’t replicate.

Of course, Google’s 75% figure raises a deeper question for the entire industry: when more and more code is written by AI, who actually “owns” it? When a bug appears, is it AI’s responsibility or the engineer’s? How does Google internally assign accountability?

I suspect the legal world will find this harder to solve than the engineering world.

The bottom line: 75% tells us that AI coding has moved from “trend” to “reality.” But reality is complicated — it’s not a binary “replaced vs. not replaced” situation. It’s a quiet restructuring of responsibilities.

Are you ready to redefine your own value?