Dennis I. Barrett programming • math • data

Coding on Copilot

💡 This is a Data Dispatch post.

Coding on Copilot, a recent report by GitClear, shows “downward pressure” in code quality over the past four years 🤖📉.

They find increased addition of new code, significantly decreased movement of code (which suggests less refactoring), increases in copy/pasted code, and increases in churn code (i.e., code changed within two weeks).

These trends are suggestive of the effects of LLM coding tools like Github Copilot, which (as I’m sure many of you can attest to) can make writing code much faster, but with the temptation of not properly checking if the code is correct.

The result can be decreased code quality, as well as extra “fixes” having to be implemented shortly after authoring code. Since these tools aren’t going away, I think the takeaway should be that we need to find improved ways to prevent this decrease in code quality.

Code reviews are an obvious foil, but I think there’s also scope for new tools (maybe even LLM-based) for refactoring and improving code quality.

Find the whitepaper here: