I built an entire web platform (Node.js backend, vanilla JS frontend, Supabase database) where roughly 80% of the code was AI-generated. I have been maintaining it for about two months now. Here is what actually surprised me. The code works. That was the first surprise. The AI produced functional, deployable code faster than I could have written it myself by an order of magnitude. But "works" and "maintainable" turned out to be very different things. The biggest issue is what I call **style drift**. Every time I asked the AI to add a feature, it would solve the problem slightly differently than the last time. Different error handling patterns in different files. Inconsistent naming. One API route returns `{ error: "..." }`, another returns `{ message: "..." }`. None of these are bugs, but together they create a codebase that feels like it was written by a team of 15 people who never talked to each other. Debugging is harder than expected, but not for the reason people assume. The code is not "unreadable" -- it is actually quite clean line by line. The problem is I do not have the mental model of *why* it was written that way. When I wrote code myself, I remember the tradeoffs. With AI code, I am essentially reading someone else's codebase from scratch every time I need to fix something. The sneakiest problem: **the AI is confidently consistent about its mistakes**. It introduced a subtle encoding bug in one endpoint, and when I asked it to build five more endpoints, it faithfully replicated the same bug in all of them. I did not catch it until users reported garbled text. Fixing it meant touching every single route. What I would do differently: establish strict patterns and conventions BEFORE generating code, not after. Give the AI a style guide and architectural rules. The productivity gain is real -- I shipped a product I could not have built alone -- but the maintenance tax is also real, and it compounds.