A Quarter of Google's Code Is AI. Who Reviews It?
On Google’s earnings call this week, Sundar Pichai said that more than a quarter of all new code at Google is now generated by AI, then reviewed and accepted by engineers.
Read that again. Google. One of the most demanding engineering organizations in the world, with some of the most complex systems ever built. A quarter of the new code.
That sentence has two halves, and I think the second half is the important one.
The first half: AI writes the code
Let’s appreciate how fast this happened. Two and a half years ago, ChatGPT didn’t exist. Copilot was a novelty that suggested the next line. Now a quarter of the code at Google comes from AI, and if you ask around, the number at many smaller, faster-moving companies is higher.
I’ve written this year about how AI turns senior developers into 10x developers and how the core activity is shifting from typing to reviewing. Google’s number is the strongest public evidence yet. This isn’t a demo, and it isn’t a startup’s pitch. It’s the reality of how production software gets written at scale, in 2024.
And this month, two of the people behind modern AI - Geoffrey Hinton and Demis Hassabis - won Nobel prizes, in physics and chemistry respectively. The technology that’s writing a quarter of Google’s code is also folding proteins. We are living through something historic, and it’s easy to lose sight of that in the daily noise.
The second half: engineers review it
“Then reviewed and accepted by engineers.”
This is the part that deserves more attention. If AI writes a quarter of the code, then a quarter of the code is code that no human wrote, which means its quality depends entirely on how well a human reads it.
Reviewing code is a different skill from writing it, and in some ways a harder one. When you write code, you understand it because you built it, line by line. When you review code, you have to reconstruct the intent, check it against the system, imagine the cases it doesn’t handle, and notice what isn’t there. Doing that well, quickly, at volume, requires deep familiarity with the codebase and long experience with how software fails.
The bottleneck in software is moving from writing to reviewing. And reviewing is senior work.
What this means for the industry
If you follow the logic, a few things become clear:
Review capacity is the new constraint. A team’s speed will increasingly be limited not by how fast it can produce code, but by how fast it can verify code. Teams that are good at review will move faster. Teams that rubber-stamp AI output will accumulate bugs and debt at a speed we’ve never seen before.
Experience compounds. The developer who has seen a thousand bugs recognizes the thousand-and-first. AI doesn’t change that; it gives that developer more code to apply it to. This is the K-shaped market in action.
Simplicity matters more than ever. Code you can review quickly is code that’s simple, consistent, and well-structured. A messy codebase makes every AI change harder to verify. Boring architecture is a review-speed multiplier.
Tests and types become guardrails, not paperwork. When you didn’t write the code, automated checks are what tell you the AI didn’t break something you weren’t looking at. Invest in them where it matters.
For developers
If a quarter of your job is going to be reviewing AI code - and soon it’ll be more - get deliberately good at it:
- Read more code than you write. Other people’s, open source, the AI’s.
- Learn the classic failure modes: off-by-one, null handling, race conditions, injection, N+1 queries, timezone bugs. These are exactly what slip past a quick glance.
- Ask “what’s missing?” as often as “what’s wrong?”
- Don’t accept code you don’t understand. If you can’t explain it, you can’t own it.
For leaders
Your senior engineers just became your most important quality control layer. Protect their attention. Don’t measure them by lines of code - a senior developer who rejects a bad AI change has just saved you more than one who merged ten good ones.
AI writes the code. Seniors decide what ships.