In this sense, anything done with human experts isn't really "clean room" either.
Suppose you're trying to reverse-engineer some proprietary tool, for instance. If you're doing that with human experts, you pick experts who haven't seen the code of the tool, or any disassembly of the tool, or any potentially tainted non-clean-room analysis of the tool.
Even if you have two separate AI instances do the RE and the reimplementation (which most people don't), the AI you use may well have been trained on the proprietary source code, or on any number of disassemblies/analyses/reimplementations/etc that aren't clean-room.
That's doubly true when people do ridiculous things like AI reimplementations of publicly available GPLed source code, where the AI definitely was trained on the GPLed implementation.
So with people, much like LLM, there's a good chance they saw the same code and papers the implementers of your target did, and whether your work is clean-room or not boils down to believing or disproving they were not aware of the association.