Does AI make code quality worse?

It can, in a measurable pattern. GitClear's analysis of 211 million lines found duplication rising and refactoring falling as AI adoption grew. AI doesn't write unreadable code; it nudges teams toward pasting variations instead of consolidating, which raises maintenance cost over time rather than at commit.

What did GitClear find about AI and code quality?

Across 211 million lines over five years, GitClear found copy-pasted lines rose from 8.3% (2020) to 12.3% (2024) and, for the first time, exceeded refactored ("moved") code, which fell from 24.1% to 9.5%. Code clones grew roughly fourfold, and the share of code revised within two weeks of commit climbed from 3.1% to 5.7%.

Why does AI lead to more duplicated code?

Because generating a fresh block is now cheaper than finding and reusing the existing one. Refactoring requires understanding the whole codebase; pasting a new variant doesn't. AI lowers the cost of writing more code, so the path of least resistance, duplication, gets taken more often.

How do you keep code quality up while using AI?

Treat AI output as a draft, not a final answer. Require review that consolidates duplication instead of waving it through, run clone detection in CI, and track duplication and two-week churn as first-class metrics. The goal is to spend some of the speed AI gives you on the refactoring it skips.

Does AI hurt code quality? GitClear's 211M-line answer

tsukumo

Does AI hurt code quality? GitClear's 211M-line answer · tsukumo

Two paths for the same change

Criterion	Paste a new variant	Refactor and reuse
Cost to write	Seconds	An afternoon
Needs whole-system context	No	Yes
Passes the test today	Yes	Yes
Cost of the next change	Edit in 3 places	Edit in 1
What the metric shows	Velocity up	Velocity flat

AI ships code by copy-paste. GitClear measured the bill.

What GitClear measured#

Why AI pastes instead of refactors#

The bill is deferred, not avoided#

What the data does and doesn't say#

What keeps quality up#

What to do on Monday#

How we think about it#

How we run a 9-agent growth team on wrai.th (and what broke)

AI 'reasoning' has a cliff. Apple went and found the edge.

Your multi-agent system isn't failing on the model. Berkeley counted where.

Want this running on your team?