On reviewing and interacting with LLM generated code

January 12, 2026

Lately I have been interacting a lot of LLM generated code. Here are some observations of how I'm experiencing the integration of this tools into our day-to-day lives (as a programmer that needs to review that code).

LLMs for Pull Requests descriptions

  1. They tend to be misleading as they assert things that do not happen in the code changes.
  2. Are super generic and too verbose. I know how to read code so I do not need an extra 500 words of comments for a function change. If I need to review 5 or 6 PR a day, I do not have time to read 5 pages of generic AI text.
  3. Don't provide any insight on why the developer made the changes, why are they needed, what's the benefit, etc. It is just a description of the change.

LLM Generated Code

  1. It is extremely obvious that it has been generated by an LLM. The generated code lacks of context and usually they do not follow conventions, patterns or style of the codebase. Is like adding to a book a chapter written by another author with another style. It just doesn't fit and creates noise.
  2. I think that programmers think they are empowered and therefore try to bite more than they can chew. I'm seeing more "try to do everything" kind of PR which is a bad pattern.
  3. Programmers over design solutions because LLMs "can handle that". This has a huge impact on the Product Design, you end up with bloated libraries and codebases just because LLMs can generate code. However: are those methods useful? do they provide a nice API? Do they return useful values to other developers? I'm finding codebases that do a lot of useless things, they do everything except the important stuff.
  4. It is somehow killing critical thinking. Submitters cannot answer simple questions like: why did you implement this change? They just assume that if the LLM wrote the code it is for a reason.
  5. I HATE the verbosity of comments every 3 or 4 lines of code. I don't want to read what you are trying to do but what are you actually doing. Again, I know how to read code, do not add useless translations to it. At the end it makes code more difficult to review because it miss leads all the time. I have seen plenty of code directly contradicting the comment in the previous line. Even more when you ask for changes to the PR and comments start getting obsolete.

Github UX

  1. It is being completely ruin.
  2. Copilot reviews are just pure obvious things, they never provided my with a "PhD Level" review. I just found them useful for trivial things like typos of styling.
  3. Github is ruining the platform with SO MUCH AI in it. I recently had to review a PR that was "already reviewed" by Copilot and the ammout of useless comments that I need to close/hide/minimize in order to actually read the code made the UX just more complicated.