over 25% of Google's new code is now written by AI

Sundar Pichai just revealed a wild stat on Google’s Q3 2024 earnings call: more than 25% of all new code at Google is now generated by AI, then reviewed and accepted by human engineers.

Basically, the company that builds software for billions of people is now letting machines write a quarter of its own software.

The internal AI named Goose

Pichai said AI-assisted coding is boosting productivity and helping Google engineers "do more and move faster."

This didn't happen overnight. Business Insider reported back in February that Google deployed an internal AI model called "Goose" to help staff write code and build tools. Internal documents revealed that Goose was trained on 25 years of Google's internal engineering expertise.

While leadership has previously reassured staff that AI isn't taking their jobs yet, crossing the 25% threshold is definitely turning heads internally. Some employees are left wondering if they are effectively automating themselves out of a paycheck.

Big profits, but a shadow of a breakup

AI is also juicing Alphabet's bottom line. The company pulled in $88.3 billion in revenue this quarter.

Google Cloud, which sells AI infrastructure, surged 35% year-over-year to $11.4 billion. Even wilder, its operating income jumped from $270 million last year to $1.95 billion. Meanwhile, YouTube's combined ad and subscription revenue topped $50 billion over the past four quarters for the first time.

Everything looks stellar on paper, except for the massive regulatory storm overhead. Following an August court ruling that declared Google an illegal monopolist in search and advertising, the US Department of Justice is weighing remedies, and a forced breakup of Google is officially on the table.

Why it matters

We think this 25% figure marks a psychological turning point for software engineering.

For developers, the debate over whether AI can write production code is over. If Google's high-standard monorepo is already running on a quarter machine-generated syntax, relying on AI coding companions is no longer a luxury. It is becoming the bare minimum baseline for developer throughput.

For the tech industry, Google is showing that proprietary internal data is the real secret sauce. Training an internal model like Goose on 25 years of institutional code conventions gives them an efficiency edge that off-the-shelf LLMs cannot touch. Expect every enterprise to copy this playbook.

However, there is an uncomfortable tradeoff. Having engineers act primarily as reviewers rather than authors sounds great for speed, but rubber-stamping code is notoriously draining. If junior engineers spend less time writing architecture from scratch and more time reviewing AI drafts, we have to wonder whether institutional engineering intuition will slowly begin to erode.