AI or not to AI......

AI, AI everywhere, in my inbox, in my news feeds, in my conversations, in my podcasts, in my meetings. If you're in tech, it's flooding your brain, if you're not in tech, well, it'll be flooding your brain soon if it hasn't already.

If you're a developer like me? Throw me a life jacket, I'm starting to drown.

BUT am I drowning in delight? Not quite. 

I was doing some work on a large legacy codebase. A member of the development team, I'll call him Jim (anonymised), wanted me to review his latest branch. Jim had been working on it for many weeks, and felt it was ready for prime time. I've been a senior for many years, and I still look forward to developer prime time.

Aha, a bug fix, nice and easy. But... why had large areas of the file that contained the bug been rewritten? And just for a minor UI issue? I asked Jim, and he said, "If you let me know which areas of the code you don't like, I'll rewrite them". The problem was it was most of it. And that's a tough response to give any developer, particularly someone early in their career. Though the code may have  worked (I didn't test it), it just didn't make sense why the file had been modified so heavily, and in ways that I wouldn't even expect from a professional coder. I fed this back. Asking "did you use AI to write this?".

The challenge we face as developers is the need, the desire, to write good code. But good code isn't just code that compiles. It isn't just code that runs. Good code is code that makes sense. Good code can be followed and understood just by reading it. It is code that fits seamlessly into the codebase such that one file has the same conventions as any other. If I can understand and follow one, I can understand and follow any.

But current AI tools don’t understand this in the way humans do.  It doesn’t understand coding in context.  It doesn't inherently know about conventions, either implicit or explicit. It doesn't know whether the code it produces compiles, runs or works within a specific codebase. It has no concept of "compiles", "runs" or "works.  And yet here was Jim, wanting to commit AI-written code to the codebase, to sit in Git for all time, as if it was a done deal.

The use of AI here was masking a gap in problem-solving and code comprehension.  My five-line solution that left the original file broadly intact came from years of domain experience, decades of problem solving on a broad range of codebases both large and small. A human database of knowledge about not only code, but contexts, requirements, gotchas, coding styles, and in-house ways of working.

These are things that, today, can't be replaced with AI. But maybe I'm being elitist? Maybe AI code is "good enough"? Unfortunately, I've worked on codebases full of "good enough" code, and when you've spent literally weeks debugging for every one feature added, you start to realise that "good enough" really is not "good enough" at all, and both time and costs can spiral out of control.

Do I think AI will get better? I have no idea. Until it can understand codebases and context, it's a far stretch. Does it have its uses? Definitely. But for every astounding sales pitch, we must, as professionals, remind ourselves of some uncomfortable facts:

“Linus Torvalds is OK with vibe coding as long as it's not used for anything that matters” https://www.theregister.com/2025/11/18/linus_torvalds_vibe_coding/?td=rt-3a

"AI coders think they're 20% faster, but they're actually 19% slower" https://youtu.be/96j1cIMYV1U?si=FU4gcOCC9G4v1ssw

"Cursor CEO warns vibe coding builds ‘shaky foundations’ and eventually ‘things start to crumble’"

https://fortune.com/2025/12/25/cursor-ceo-michael-truell-vibe-coding-warning-generative-ai-assistant

And this is why, right now, I do not use AI to write my code. Yes, I'm fortunate, given my long and winding history, I have a deep enough skillset to not need to. I am also genuinely concerned my skills will diminish if I do. I also surmise I may even be slower if I do. No, this might not be true for everyone, but my successes paint a picture that's hard to ignore, and if developers want to reach these heady heights, leaning too hard on AI right now may not be the best plan of action.

Disclaimer:

This post reflects my personal experience and views as a software developer working with large, long-lived codebases. It is not a blanket criticism of AI tools or the people who use them. Experiences with AI-assisted coding will vary by context, team, and level of experience.