AI IDs should not be eligible for Community ID Opt-outs

If the ID is based solely on the AI’s all-too-often incorrect assertions, there is zero reason that ID should be able to opt-out of the community ID. For inattentive users who do this, it can create misidentifications that are very difficult to correct, particularly in taxa with low engagement on here.

I agree that this would help the issue, but there’s a problem: it’s impossible to tell if an AI suggestion was accepted because the user didn’t know, or because it was faster than looking up the name of the taxon they wanted.

14 Likes

I was thinking about this today. Is there a way to prompt/require the IDer to write even a few short words saying why they select the suggestion? A way to link tag words to the iNat ‘About’ description? Most IDers seem to just click and go. … I posted my observation of a very minimally iNat observed lifeform today, and I can see where this type of information gathering is really lacking in these cases. … I’m not sure I’m explaining my thoughts perfectly here, but I just think a little more effort into quality IDs would definitely be a good thing for members and for the AI.

Implementations like this rarely produce value. They get seen as an unnecessary slowdown in entering data. People will just hit the space bar, enter sarcastic stuff like ‘because that’s what it is’ etc

11 Likes

I agree with the sentiment of the feature request, but I think you’re right on @aisti. There’s no way that I can think of to tell if an ID is solely based on AI. Heck, I often just click and let the AI populate because it’s quicker than me typing some long scientific name that I’ll probably mess up a keystroke for…:)

I do usually note in the description if my ID is solely based on the AI (hopefully to encourage people to contradict it if it is wrong/bad!) which I think is a good practice, but can’t obviously be required.

7 Likes

To my mind the AI ID is meant to act as a starting point, not as a definitive identification.

It helps to get an observation to a place where others actually see it and then correct or confirm the ID.

Without the AI suggestion many IDs would wind up being left in the “life” limbo and never seen by people who know about that particular branch of life and who can move the identification in the right direction.

Yes, the AI does make a number of mistakes, but it’s meant to be a suggestion. Where I am, in tropical Asia, the AI has a lot of trouble with species, but it’s still very useful and even placing an incorrect ID from the AI winds up being far more useful than leaving it blank or making a stab in the dark. When the initial ID, AI or otherwise, is wrong it’s an easy matter to simply withdraw that identification.

One thing that might get at your concern while not suggesting an unworkable change to the system is to have an ID that’s purely AI based marked as “Preliminary ID”. The the observer would have the option of independently confirming it in a separate field, just as if another person had made the ID.

4 Likes

Could someone explain when it is actually useful to opt out of community ID, because the only times that I have seen it used is when someone is very wrong in their ID and usually just took the suggestion. So I liked the change a while back switching the observation to casual if the community disagrees, at least getting it off the maps. I like the idea of making it so you can’t opt out if you are taking the AI suggestion.

1 Like

I live in South Korea and have periodically had people come along and identify my observations as species from the USA despite there being multiple similar-looking species here in East Asia. When I ask about what characteristics lead them to their ID several users have said that they didn’t look at the geographic location when submitting their ID but others haven’t responded at all despite me tagging them a couple of times and those users still being active on iNaturalist.

If it’s my genus level ID and two species level IDs from another continent I mark ‘ID can be improved’ in the DQA section. If it’s my species level ID plus a trusted local user’s species ID and two+ species level IDs from another continent I do sometimes think about opting out of the community ID. (I think I’ve only actually done it once or twice though …)

9 Likes

The only time I’ve done it is when someone identified my photo of a grebe as “birds” on the basis that they couldn’t be certain it was a grebe from the photo. I thought they were probably right so I opted out of community ID and marked as “id can not be improved” which made it casual grade. I was certain it was a grebe as I saw it closer or through binoculars. I reckon a casual grade observation has more value than leaving something which may be unidentifiable in needs id.

You don’t need to type the whole thing. If you remember the first 3 letter of the genus & species that’s usually enough (e.g. “hom sap” will give you human). The same works for common names too (“ha to fe” willl give you Hart’s Tongue Fern). I can’t remember where I learnt this but I’m quite grateful as I can’t remember how to spell many scientific names, let alone type them correctly.

6 Likes

On the face of it, this seems a dubious conclusion. Surely it’s also possible that the AI suggestion is right/plausible, and the current Community ID is wrong/questionable? In which case, why should a user be prevented from opting out just because they used the AI to add their ID?

4 Likes

I agree. This is one of those ‘free will’ issues that never get resolved between humans and AI. … Lol … I am very aware that it’s rather idealistic to think that some others would see the big picture value of trying to improve the programs identification capabilities. … And of course those are the others who are quickest to complain when it doesn’t work like they want it to.

1 Like

Actually the computer vision’s top suggestion (“We’re pretty sure this is…”) does a lot better than most, users. The problems is that those users that have no idea what they are looking at, instead of selecting the suggestion “We’re pretty sure this is…”, simply select the photo that most closely resembles the observation. And that’s not even a photo - it’s a tiny thumbnail. So basically they select something that has the same color instead of selecting the best suggestion of the computer vision. Obviously the thumbnail of a genus can’t illustrate even one, let alone the (say) 100 species that it includes, and most often the species in the thumbnail does not resemble the one in the observation. So users will select a specific species further down the menu that has the same color. And most often it is incorrect. Obviously only one species can be correct, so a list that includes 10 species will have either 9 or 10 wrong species. So the unknowledgeable user has a high chance of choosing an incorrect suggestion. Thus, it’s not the computer vision’s fault - which most often is correct with its top suggestion - but it’s the user’s fault that they do not select the top suggestion when they have no idea what they are looking at.

How do I know when they don’t select the top suggestion - if it’s a species, it’s not the top suggestion which is never a species. Also, I check the suggestions afterwards and check if it was the top one. But that only works soon after the ID has been made because the computer vision models changes with time and so do the suggestions. Thus, it would be nice to be able to see if the user has selected the top suggestion or some other from the list. It can be helpful for user behavior analysis.

I also pick up AI suggestions because it’s easier than typing and because I get a second opinion. And sometimes I pick the top suggestion (not uncritically though) when I have no idea what I’m looking at and the observation has no ID. That’s the correct behaviour.

3 Likes

An excerpt from the iNat help page:

  1. When should I agree with someone else’s identification?

An identification confirms that you can confidently identify it yourself compared to any possible lookalikes. Please do not simply “Agree” with an ID that someone else has made without confirming that you understand how to identify that taxon. If you agree with the ID without actually knowing the taxon, it may reach Research Grade erroneously.

In effect, many (possibly most) observers who use Computer Vision suggestions are just agreeing with somebody else’s ID without understanding how to identify the subject.

My suggestion would be that rather than just selecting a taxon from a list, there would be two options:
Accept - meaning the suggestion looks good but the observer cannot independently confirm
Confirm - the observer is able to independently confirm the CV suggestion

Identifications that use the “Accept” option would not count toward the community taxon, so it would require two additional identifiers to get to Research Grade. It would also make it possible to more easily separate the IDs based only on CV from those that just use CV as a way to reduce typing.

I would also suggest that CV never make suggestions for taxa that are outside their known geographic range, make higher level suggestions rather that species when diagnostic feature are not visible, and a few other things - but these are even further off the subject than I have already gone…

1 Like

You are right and the same aplies to the IDs of the many unexperienced users who try to propose an identification without being supported by any knowledge of living organisms.
But the only way it to recruit more experienced and skilled identifiers together with suggesting to unexperienced users not to trust too much in the automatic IDs as well as providing an ID at the kingdom rank whenever it is possible.

PS: I have noticed that for the Italian flora the automatic IDs have improved considerably.

4 Likes

I think that the terms “accept” and “confirm” are perhaps a bit too similar and would lead to rampant confusion, but the premise of this idea is sound. Perhaps something more like “Provisional ID” and “Confirmed ID”, with the former only counting towards Research Grade when a confirming ID has been given. The idea might better be discussed in a separate post.

3 Likes

I like your idea very much.

It encourages participation by those who are still learning, while also improving the quality of the IDs.

1 Like

I don’t think that needs to be a barrier to implementing the proposed feature request. All the observer needs to do is withdraw their original ID and type a new ID rather than using computer vision to suggest it.

One important consideration is that we would expect the frequency of “novice users who choose opt-out of community ID without much understanding of the impact” to be a whole lot higher than the frequency of “users making informed decisions to opt-out of community ID who happen to have used AI to suggest the original identification”. And for the second scenario, those few users can fix that by removing and re-adding the ID.

Even more efficiently, when making the original ID, select the desired AI choice, then just type Space and Backspace before saving. It will then be recorded as a user-typed ID instead of an AI ID.

4 Likes

Yeah, the terminology needs to be clearer. In the unlikely event that anything like this gets implemented I would put in an initial pop up message that clearly lays out the options and their implications. And then periodically returns to remind people - with an option to suppress the reminder in one’s personal settings.

2 Likes

But that clearly is a barrier. Why should users who are using the system correctly be penalised in any way at all?

Fix what?! They used the system exactly as intended!

1 Like