I disagree that people using AI to make IDs[1] is harmless.
I occasionally get random IDs from people who seem to be using the CV to make IDs with very little additional research besides comparing images and intuitively guessing which ones look similar. I tend to react with a fair amount of annoyance for the following reasons:
If they are doing so for difficult taxa where there are few skilled IDers, it can be quite difficult to get enough IDs to override that one incorrect one if the user does not withdraw their ID – if someone sees and corrects the ID at all.
If the uninformed ID is for a species that generally cannot be identified from photos but people often try to do so anyway, it is not unusual that there is a certain amount of IDer fatigue and IDers might add a non-disagreeing genus ID instead of disagreeing, so I get stuck with a random ID that I doubt is correct but I might not know enough to assess myself beyond general experience that the taxon is difficult.
Users also have a tendency to treat IDs provided by others as having some authority – regardless of whether that other person has any knowledge or not. So they might agree with the ID that the other person provided because they assume that if someone went to the effort to provide an ID it is based on some knowledge and not a random selection of one of the CV suggestions. Again, if it is a taxon that lacks identifiers, these wrong IDs on an observation that has just become RG based on no knowledge at all might remain uncorrected for a very long time.
And finally, if someone is using the CV or AI to ID my observation, I do not learn anything from it. If I wanted to see what a computer would tell me, I am capable of doing this myself. If it is a taxon where I lack knowledge myself, I might in fact have decided not to use the CV suggestions because I can’t evaluate them and I know the taxon is somewhat tricky and I want feedback from someone who knows more than I do. If they are using the CV or some other AI to make their IDs, I am getting uninformed suggestions masquerading as information. iNat’s validation system only works if we can rely on people making IDs with care and thought.
By “using AI to make IDs” I mean relying on it to make IDs that one cannot evaluate oneself. Obviously there are other ways of using AI tools – the key word here being as a tool, namely to brainstorm ideas or help one figure out where to start looking, followed, crucially, by critical assessment of those suggestions and additional further research of one’s own. ↩︎
I never said anything about enforcing a rule that people cannot use the CV in any way.
But we can encourage a culture in which it is not considered acceptable to rely exclusively on the CV to ID other people’s observations[1] without drawing on additional sources/knowledge.
In the Identify module the CV is not forced on anyone, nor is anyone forced to add identifications to observations of taxa they lack knowledge about.
I realize that people are likely to continue to use it fairly uncritically to ID their own observations – that is their choice, as much as I wish they wouldn’t. But forcing uninformed IDs on other people who may not want them has far-reaching negative effects on the community. Yes, I realize that observers can opt out etc. in such cases, but it is a poor fix for a problem that should not happen in the first place. ↩︎
I have encountered one recent case out of my thousands of IDs as this (but ofc this AI IDing is obviously going to blow up and gonna cause trouble for me as IDer soon - I have raised issue of people using AI notes to observations before but no inat stance there yet)
On that case, I have given a most recent phylogenetic paper showing clearly why their ID is wrong and even how the exact observation location is completely out of scope of their ID when adding my disagreeing ID.
I got hit twice from observer and another IDer who seems to be friend to observer with an AI reasoning lengthier than an essay with such an ironic way that a 5th grader can catch on. Basically, it found some expert author names from somewhere and it found locations from elsewhere and it mishmashed things serving as a super tight reasoning while its entire reasoning relying on a faulty premise would break things if one reads those text again. Basically I can see how if a question asked to AI as “Prove me why this is species X and not species Y that this other IDer is saying with comment Z” would definitely come up with such contorted reply with such blatant loopholes in reasoning.
when I called out why its chatgpt reply and how it is making faulty reasoning to things I already showed and linked by even contorting the real geography itself - the observation is deleted :) (atleast there are no long dialogues but maybe because they are called out early)
on another note, I have seen another observer finding species name from advanced LLM using latest research and proper unbiased dialogue to come to exact species name for a completely rare species that I am so hyped to get me tagged upon. I was searching for days everywhere to see even one single observation for that entirely new Genus, but since the genus itself is new and even non-existent on inat, I cant even find those observations easily to ID directly too.
but in this case, the observer used AI in clever way to arrive at a reasonable guess that no other CV or AI or non expert can do, and then tagged me to confirm if that is really the species and not lookalikes. Note the all common “tag experts only and never use AI” is also not going to work always as in here when the genus is itself new and its not even easy to say the family directly too as there are lookalike families. aka finding who are even the relevant experts to that photo is itself an hard problem too in some cases that is not easy to tackle always especially if there is not enough downward specificity push in IDs that causes those observations to be in taxa experts inat URL filters. A completely rare species sitting at kingdom or such for weeks at obscure location is rarely gonna get a sudden expert species ID although there are lot of taxa experts who wade through those backlogs (i do sometimes)