Could AI code review pipelines make code quality worse?

I’m worried that adding AI to our code review pipelines might actually make our code quality worse - not because of the AI, but because of us.

I was asked at work today if I tried to integrate any AI code review tools in our pipelines already. Actually I haven’t yet, but I was thinking about it quite a lot. I did ask my local Claude Code to review a few merge requests (that’s what they are called in GitLab), to get some idea of the process and to compare the results with my own notes.

As my micro-experiment showed, Claude was really good at catching little flaws. It wasn’t that successful with some of the larger architectural concerns, which was also expected - the relevant knowledge and the bigger picture were mostly in my head by that time, not anywhere in the context of the agent.

But I actually wanted to talk about something else here. It is very tempting to configure AI code review, and it’s quite simple to do - there are plenty of tools already, even Claude Code itself can be easily run in a CI pipeline. However, I’m worried that by adding it, it is very easy to unintentionally push a human out of the loop. Like as a code reviewer, when you see that the review was already done by someone else, you may naturally pay less attention to the details. I do not justify this attitude, but I’m afraid that’s how it works for a lot of people. By doing that, we remove part of the responsibility for a thorough review from a developer. Review of the code which is already mostly written by LLM.

In addition, AI reviewers make a lot of assumptions that are untrue, highly opinionated, or at the very least, debatable. It creates friction that is not always optimal during the code review, as in general it should happen earlier, during feature technical planning.

Don’t get me wrong, one agent reviewing another’s results of work is very useful. Especially with different AI models. I do it all the time when I brainstorm, plan, write code or docs. I just haven’t decided for myself yet how to introduce it in the review pipeline, without the side-effects described above.

Does anyone have suggestions?

This article was originally published on LinkedIn on 2026-01-28.