So one option would be to comment placeholder text along with adding a broad taxonomic ID.
I’m an admin for a project collecting angiosperms, and we have a corresponding collection project collecting unknowns in the target countries. I’ve written a macro program which can add placeholder ID as a comment, and then looks at the visually similar suggestions for the observation, pulling down the top 10. It then applies taxonomic filters (‘Animalia’, ‘Fungi’, ‘Plantae’, ‘Angiospermae’) in turn and pulls down the top 10 visually similar suggestions for each in a loop. If all of the top 10 match the top 10 with no taxonomic filters, it assumes that filter is correct. e.g. if Animals match up then observation is an animal; if plants match then check angiosperms and if angiosperms also match then is an angiosperm, if not then it is a plant. It can then add that ID (the broad taxonomic filter which was correct) with a comment saying it is a broad taxonomic ID based on iNaturalist suggestions.
Philosophically this feels alright to me - we’re retaining placeholder ID as a comment and are increasing the likelihood that a user gets their observation ID’d by increasing the chance that someone will see it. It’s not ID by AI as much as increased chance of human ID thanks to AI.
Concerns about this approach:
If multiple taxa are present in a photo it doesn’t consider which the user may want ID’d (but then the observer can always add their own ID e.g. to plant if I’ve ID’d it as an animal).
If there are issues with the observation (e.g. observer has submitted multiple photos of completely different organisms), it would rely on this having been detected by different taxonomic suggestions. - suggested fix: only suggest IDs for observations lacking any comments (under the assumption that a kind user will have commented asking the observer to delete the observation and repost as multiple separate observations).
I should stress that I have not run this at all yet, and will not do so until I have some feedback from the community as to whether you think this is (a) a good idea, or (b) have any further checks you’d like to see implemented before it is run. When I do run it I would be testing on older observations in the unknown collection project I’m admining.