I agree, the policy should be to not allow low quality submissions and have clear copyright status. There are some nuances that needs to be addressed though.

With LLMs it's too easy to replicate implementation of algorithm that is copyrighted, with manual coding it's more deliberate. Sometimes you can't really separate the algorithm from the implementation. So a maintainer needs to check the origin of any algorithm used.

The copyright status for non-algorithm code is quite straightforward and it's often a non-copyrightable code anyway (boilerplate, configuration, obvious steps, etc.). The generated code by LLM has no human author therefore it is public domain, there is no problem of incorporating of public domain code. Any copyrigtable modifications are made by the user of the LLM.

The biggest problem is the amount of low effort contributions caused by AI usage. But one could just reject it on being of low effort.

There is no need to specifically target AI in the policy, with one exception that is unique to LLM contributions: people not understanding anything about the produced code at all. One couldn't trust anything from such contributor.