But also, from personal experience with the output, at a very mundane level it destroy comprehensibility of the code base and documentation. People will use LLMs to generate documentation that is voluminous but also is targeted towards other LLMs rather than human comprehension.
The net result is that even if a stated rule is that people will review and understand the code, that becomes functionally unachievable. This becomes doubly problematic when grappling with generated documentation for integrating components in a larger system. The code changes downstream then become larger than they needed to be because the documentation was more verbose and less comprehensible than it needed to be.
These are negative outcomes that I have seen in action.