GitClear white paper on effects of AI coding assistants. Their takeaways mirror some of the early feedback I’ve gotten: - AI tools help you write more code faster - AI suggested code has a higher churn rate, suggesting authors release it and later learn it’s not working out

A more surprising concern: - AI suggested code is correlated with less code refactored and more code copy/pasted

I first considered that more code churn is OK if AI tools can make up the productivity difference. After all, failing fast and improving is at the heart of agile.

The last point about less refactoring, more copy/paste changes gives me pause. Reading code is much harder than writing, and it’s a substantial part of our job. I would extend that to say reading code gets harder at a faster rate than writing as code volume grows. With a low barrier to avoiding refactoring by using AI suggestions, we may make it increasingly harder for future engineers to interpret old code, and do so at a rate that AI tooling can’t help us with.

This said, I love copilot. I think it can quickly help you find your way around and reduce keystrokes. I think you need a skilled author at the helm to sniff when something is wrong and course correct.

Q: should GitHub Copilot be a privilege for proven engineers or used by all?

https://gitclear-public.s3.us-west-2.amazonaws.com/Coding-on-Copilot-2024-Developer-Research.pdf