When George Hotz described AI-generated code as "slop," everyone knew exactly what he meant.
His blog post, "The Eternal Sloptember," names something that's been nagging at every developer who's spent real time with AI coding tools. The code appears correct. It performs correctly. And then you ship it, and something quietly breaks in a way that takes you three hours to trace.
According to Hotz, AI generates "statistically plausible but subtly broken code." This encapsulates exactly how I've sensed things for a while now. The AI almost nails the solution. It almost gets the imports correct, the function signatures correct, the overall flow correct.
However, the final 10% that includes edge cases, error handling, strange interactions between your auth middleware, and that one legacy endpoint, is not as successful. Yet, that final 10% is the most important part.
Here is something AI is actually really good at. It can get you from zero to "looks like it works" faster than any tool previously created.
→ Scaffolding a new feature? Incredible.
→ Writing boilerplate CRUD routes? Chef's kiss.
→ Generating test stubs? Solid starting point.
The first 90% of a task feels like magic. You're flying. You feel like a 10x engineer. Then you hit the wall — the part where the code needs to actually be correct, not just plausible. That's where AI taps out and you're back to being a regular human reading stack traces.
This is where I depart from the "Eternal Sloptember" narrative. Hotz's argument suggests that AI code is inherently sloppy. I'd argue that developers are simply erroneously treating the AI's writing as final code, rather than a rough draft.
A first draft you don't review is slop. A first draft you refine is a workflow advantage. The distinction is entirely on the human side.
The growing concern — that AI empowers low-performing devs to generate 10x more low-quality code — is real. However, this is not an AI issue. It's a code review issue. It's a team standards issue. We've always had bad developers writing bad code. They were just writing it by hand. 🤷
The scariest part isn't bad code. It's the confidence bad code ships with.
When you struggle through writing something yourself, you have a mental model of where it's fragile. You know which parts you're unsure about. When AI writes it for you, that mental model doesn't exist. You merge code you can't fully reason about, and you don't even know what you don't know.
That's the actual problem. Not the syntax. It's the lack of understanding.
Hotz makes a good point, but the reality is that AI code can be sloppy if you allow it to be. The fact is that the tool is exceptionally good at getting you to the 90-yard line. However, software engineering has never been about getting to the 90-yard line. It has always been about the last 10 yards.
The developers who thrive with AI will be the ones who treat it like a fast but careless junior engineer: useful, productive, and absolutely not to be trusted without review.
So here's my question: have you changed your code review process since your team started using AI tools? Or are you reviewing AI-generated code the same way you review human-written code?