A study of 3.52 million changes: AI-written code passes the tests and ships to production, but costs more in review and resources

📉 A study of 3.52 million changes inside a large tech company reaches an interesting conclusion: AI-written code passes the tests, ships to production and gets reverted less often than human code. But it costs more.

A year of C++ code tracked from writing to production

For a year, the researchers tracked AI-generated C++ code from the moment it was written to the moment it shipped. By March 2026, more than two out of every three lines with known provenance were written by a model.

The cost shows up later.

What costs more, for the same amount of work

For the same amount of work:

• 1.92× more review threads that block the change
• 1.39× more comments
• nearly 2× the loops in the code
• 30-40% fewer calls to standard libraries
• +5% compute
• +8% memory

The numbers come from the paper Characterizing the Quality Profile of AI-Generated C++ in Production, published on arXiv on 6 August 2026 by Google researchers: 3.52 million changes between April 2025 and April 2026, revert rate of AI code around 0.9× compared to human code, share of AI code with known provenance up from 28.99% to 68.62%. LeadDev covered the study on 20 August 2026 with the same figures.

The point the researchers make is this: writing code was already a relatively cheap part of the job. Review, maintenance and efficiency cost far more.

AI has mostly accelerated the first one.

Four things to do when a model writes much of your code

💡 What this means, now that a model writes a good chunk of your team’s code:

• 1) Measure the total cost, not just how many PRs you ship.
• 2) Have the agent look for existing libraries and helpers before it writes new code.
• 3) Give it the architectural context, not just the task.
• 4) In review, focus on logic, duplication and unnecessary detours.

How I flipped my time between architecture and code

In my own work, since I work alone, I changed my approach completely: today I spend around 80% of my time planning architecture and features at product level, and 20% having AI write code. It used to be the other way around.

It’s the same balance I apply in custom software development projects for companies: more time on the architecture, less on writing the code.

I’m curious how you work: freelancers, developers in a team, or people doing vibe coding. The approach changes a lot depending on the project and the size of the team.

How do you organize it? Let me know in the comments 👇🏻