Interesting observation re:model assisted code - vibe coded software exhibits the average anti-patterns of the languages culture and ecosystem. .NET vibe code trends towards enterprise over/no-design (services, managers et al) JavaScript vibe code is often spaghetti/full of mixed concerns
A technique based on statistical averages of input data will - when it works - recreate average code. This shouldn't be a huge surprise. it follows. I bet it produces tests that have lots of mocks and coupling too. This too is very mid in the .NET world.
Yep, exactly that - it's not a surprise - but I think what people maybe haven't engaged with is that it's also reproducing cultural norms, not only "average code". It's encoding subtext - in the same way that ML encodes other forms of bias.
Good software design all exists in context, so while the tools are powerful, the true vibe coded "just go with it" approach will yield mediocre to bad results frequently without authorial control. Incredible tools for experts that can provide that additional literacy...
... but each ecosystem anti-pattern is amplified by default. I love these tools, but if you're not re-working and co-authoring the outputs with refactoring iterations, you'll end up in a bad place.
A models output is an exaggerated version of a language or communities anti-patterns by default. Like any programmer with just enough knowledge and not enough context.
@robbowley.net has some interesting words on this blog.robbowley.net/2025/08/08/i...
I've seen similar in AI Generated React code. UseEffects everywhere (a common React anti pattern). That is of course on top of the - standard across all languages - anti pattern of deeply nested conditional logic and duplication