Code AI can't write? - I have 6 examples

I was recently at a software development conference where there was a heavy emphasis on getting developers to use AI to write more code, rather than them typing it.

As part of the Q&A, I asked if the panel thought there were any circumstances where AI shouldn't be used to write code or where it wasn't appropriate.

They said that, as AI was being used to write the frameworks that developers are building on top of, they couldn't see a reason not to have AI write the code.

Being nervous of absolutes, and knowing how varied developers and the software they build are, I quickly came up with six scenarios where developers can't (or shouldn't) "just" point an agent at their codebase and tell it what to do:

1. Code is written and exists in an air-gapped environment (so on machines without access to the internet) for security or privacy reasons. This means Claude, ChatGPT, Copilot, or similar can't look at it. [Yes, I acknowledge that as an alternative, you may be able to use an agent with a local model, but you won't have access to the latest tools and may need very high-spec (so expensive) machines to run those models on. Time spent managing and maintaining such tools must also be considered.]
2. Code that you don't want the AI [company] to see. Similar to the above, but where you don't want to, or can't legally, let another company see your code or the data it uses. [Again, a local agent/model may be an option here.]
3. Code that is very hard to test in an automated way. If the code needs complex integrations with other systems that can't easily be mocked or abstracted, especially if it requires specific hardware to interact with, the AI agent is unlikely to be able to do and test everything for you. It may think it's finished, but you wouldn't accept untested code from a person, so why accept it from an AI agent?
4. Code that uses very obscure APIs. In my experience, when the AI hasn't been trained on the APIs you want (and need) to use (because there just aren't very many--or any--examples on the internet for it to learn from), then the AI can easily start hallucinating the methods it wishes existed, or it might even be reasonable to assume should exist. When this happens, it takes a person to correct what the AI has done/tried, and it can be easier to have a person do it in the first place, rather than have to fix the invalid approaches.
5. Code that requires a small and specific change. If the change is small and you know what needs changing and where, it can be quicker and easier to do it by hand rather than trying to describe what the AI should do. (And not have AI make other changes at the same time.)
6. Code you write for fun. Ignoring the challenge of finding an employer to pay you for the things you find fun; if, as a developer, you find writing code fun, then there's no reason you have to stop. It may not be how you earn a living, but that doesn't mean you can't write code by hand at other times.

I'm sure there are more. What have I missed?