Replying to this specific point that you keep bringing up, others here have already put into words the same general objections about this topic I also have.
Now to this argument, the comparison they set up in this study is completely moot for the real world - right now generative AIs are not being used to replace humans doing things like writing good and meaningful (!) text or create meaningful and valuable art, it is used to create additional text and additional images, which add on to the total energy consumption. (Or why else, as somebody in the thread brought up earlier, are the projected future energy costs for generative AI absolutely massive). Even if iNat implements the LLM like you describe, it would still always have to be verified and very likely modified by at least one real human expert, and with the caveats they list in the paper it’s very unlikely that the average energy per species (or wiki page, or whatever else) will be lower when an LLM is involved. Or in other words, the LLM uses less energy but can’t actually really perform most talks that a human could perform (as others have said, hallucinations are rampant everywhere and the whole internet is at this point already flowing over with AI-generated nonsense).