I don't think I would want to design such a study, but at a minimum arguments would have to be made ahead of time for meaningful pairings "outside-water" == "beach", all button presses would have to be recorded, and then a statistical conclusion could be reached.
> I mean, some people are perfectly fine with labeling other people NPCs.
Indeed. So whatever such a study concludes is unlikely to change people's ideas about dogs much.
This is actually a study being run - https://www.theycantalk.org/research
They are recording button presses and sequences in both controlled and open situations, and have already observed a consistent capacity to chain multiple buttons together in naming a thing, and have found patterns in the names dogs come up with (I think they're also doing cat studies.)
> I mean, some people are perfectly fine with labeling other people NPCs.
Indeed. So whatever such a study concludes is unlikely to change people's ideas about dogs much.