2025-07-30 In praise of GitHub Copilot code reviews
===================================================

.. feed-entry::
   :date: 2025-07-30

`GitHub Copilot will now do code reviews <https://github.blog/changelog/2025-04-04-copilot-code-review-now-generally-available/>`__. I’m a fan, but it turned out several friends aren’t. So I wanted to talk about why I really like them.

The key thing from my perspective is that **Copilot code reviews are not the same as human code reviews** [#ai-code-review]_. They are substitutable *to a point*, but not entirely substitutable, and that’s an important distinction.

**Copilot code reviews aren’t as good as a good human reviewer, but they’re hugely cheaper, and that low cost brings a bunch of benefits**.

**In the literal dollar sense,** `Copilot Pro costs $10/mo <https://github.com/features/copilot/plans>`__ and offers some unclear number of code reviews - but even if it were $10 a review, consider how that compares to a colleague’s hourly rate.

But it’s the implications of that low cost that are the most valuable and interesting to me:

**Sometimes there is no human reviewer**. `A current side project of mine involves porting a C codebase to Go <https://github.com/doismellburning/samoyed>`__ - I’m currently working solo, there’s nobody else who’s going to review my code - I’ll take a Copilot review as better than nothing! If I had a human reviewer, I bet they’d give me a bunch more valuable feedback. But I don’t.

**Even when there’s a human reviewer, Copilot saves time and energy**. Copilot can review my PR before a human has had a chance, which means it can find (potential) issues sooner, even if they’re only minor. I use formatters and linters because I want computers to handle the trivial stuff so humans can focus on the more interesting bits - I see a Copilot review as much the same. `It can read a ~2000 loc diff in ~1 minute and tell me I’ve probably used the wrong exit code <https://github.com/doismellburning/samoyed/pull/73#discussion_r2229676636>`__, and I can get that fixed before a human has had a chance to even load the diff.

Crucially **I still want a good human reviewer if and where I can**, but if I can’t then this is better than nothing, and even if I can then this helps their load.

**Is it definitely better than nothing? I believe so.** Sure sometimes it suggests things that don’t feel valuable, but so do humans - my experience so far is that these are a combination of minor, matters of taste, and infrequent enough to not be a distraction.

**Copilot code reviews are no substitute for good human review. But they’re still helpful.**

And this is basically where I’m at with the current state of AI tooling: people will try to sell you on “replacing humans” because that’s enticing to the folks holding the purse strings, but that’s not the reality - but don’t disregard the value as a force-multiplier.

.. note::
   In a wonderfully serendipitous example, Copilot caught that I'd mangled the rST footnote syntax when I first pushed this post!

See Also
--------

* `Why AI will never replace human code review <https://graphite.dev/blog/ai-wont-replace-human-code-review>`_
* `RDEL #92: How do AI code reviews impact engineering teams? <https://rdel.substack.com/p/rdel-92-how-do-ai-code-reviews-impact>`_

.. [#ai-code-review]
   For "Copilot code reviews" here you can basically just read "AI/LLM code reviews", but Copilot is the tool I've used most for this.