I think the missing piece is a shared measurement standard. Token counts are not a footprint: they omit hardware manufacture, training and model refreshes, batching/caching, grid mix, cooling, and local water stress.
Some companies publish broad totals, and Google has now published a detailed inference methodology for one service, but we still lack consistent, independently auditable figures by model and workload. Until we have them, neither “AI is harmless” nor “every use is indefensible” is a scientific claim.

I would rather see pressure for disclosure and a practical rule: if a project uses AI, state its compute/carbon/water budget, the outcome it enabled, and the simpler non-AI alternative considered.