I generally use ID With AI on all of my own observations, and choose the “we’re pretty sure” taxon if there is one and if it seems right to me, even if I’m totally unfamiliar with that taxon. (I’m not an expert in any taxon, but I enjoy learning and trying to get more correctly ID’d Research Grade observations on the map.) I’m assuming this is accepted as appropriate even though I’m IDing beyond my abilities, since otherwise I would be adding a lot of “plant” and “Lepidoptera” observations which might take longer to get seen by people who can confirm the ID.
Here’s where I’m a little confused: the AI is a lot more confident now about a lot of taxa. I’ve been going back through my own older observations that are still at a broad taxon, and the AI is now suggesting more specific ID’s as “pretty sure” for a whole lot of them.
Question 1: Is it “appropriate” / “good” to go back through my old observations and add the AI ID even if I’m not sure at all, when it brings the ID to a more specific taxon, and isn’t disagreeing with previous IDers? After all, I don’t know if previous IDers were saying “this is as far as can be ID’d with this potato-quality photo” or “this is as far as my knowledge goes.”
Question 2: Is it “appropriate” / “good” to go through other people’s old observations, and add the AI ID where it makes it more specific, when it’s not disagreeing with other ID’s, even if I have no expertise?
I get the feeling that someone coming in and adding an ID has more of a vibe of “I know this to be true” rather than “I’m just trying to help move the process along.” But it’s pretty fun and satisfying to nudge these old observations toward Research Grade in this way, and it seems to get experts to see them more quickly, so I’m wondering if it’s discouraged.
Apologies if this has been discussed, but I couldn’t find another thread on it. Thanks!
I think in general it is ok to ID your own observations a little more boldly with CV suggestions, though I often only do so to genus level myself and will note when my ID is based on the CV and not any other specific expertise that I have.
I would not go to other people’s observations and just add CV IDs or agree with another CV ID. All it takes is 2 people doing this to get an RG observation based solely on the CV suggestion. If practiced widely, this can just lead to feedback loops of CV IDs training future versions of the CV. Observers can always ID their own observation with the CV if they want, but they have made a decision to use the CV how they see fit. I’ve personally had IDers just add CV IDs to my observations and asked them to stop (and have seen stories of other users doing this on the forum as well).
Problem is - if you and CV are wrong - then it needs 3 identifiers to get it to RG.
If someone agrees with your wrong ID first … then it needs 5, and those 5 identifiers are not happy.
Rather broadly right, than narrowly wrong.
Is it “appropriate” / “good” to go through other people’s old observations, and add the AI ID where it makes it more specific, when it’s not disagreeing with other ID’s, even if I have no expertise?
This might work for some very common taxa but as soon as you ID in a region with many similar endemic species I would not trust CV at all. (CV model does not know these species with few observations yet…)
It is also not recommended to ID other observations without actually having any expertise in how to differentiate species.
Using CV suggestions in your own observations on the other hand is ok. But please do a quick research before clicking on any species.
I have observed many cases now (as a frequent identifier) where people click on the wildest suggestions (they always click on species not e.g on families or genera).
If the CV suggestions do not make sense just type in the highest possible taxon you know, e.g. ‘Insecta’ for insects or if it is a fly for example ‘Diptera’.
This way there are less taxon disagreements (with expert IDs) and other expert identifiers might find this observation faster.
But I also understand that sometimes CV suggestions might be incredibly correct and helpful. It depends on who is selecting which CV suggestions…
Personally I only ID to the level I’m confident in regardless of what the CV system says.
What I will do is check the CV suggestions (via the links) and see if any of them actually match what my observation is and if there are other observations of that species in my area that match. If that matches, then I may agree with the CV, but even then I’ll usually leave it at the genus level and put the potential species in a comment. I tend to do it this way because in my area there are a lot of potentially undescribed species, species that are very similar to each other, etc.
I think it is a very bad idea to accept the CV suggestions without checking on them yourself.
I still feel new on this platform.
My best practice so far regarding own obervtions is to check-out CV.
I take the best match as starting point for conventional search using the taxon plus “confusion” or “mistake” (in my local language). In most cases i find expert forum entries giving more detailed guidance on how error prone this taxon is and then i decide if i take the species or the genus only, usually adding a link to the information source for reference.
For other user’s entries, i am more sceptical to agree or propose something specific if i am not fool-proof sure about it. CV is not replacing human risk assessment and decision for each individual case.
You can use the CV as a guide, but please don’t use it to make IDs on other’s observations with the CV without some background knowledge. Please don’t add an ID for the sole reason of “the AI says it is this”. I used to do this when I first started five years ago and it still affects my reputation to this day.
Your observations are a different story. Part of Inat is trust, so people trust that what you identified it as is actually what it is. For your old observations you can use the AI to narrow the identification down, but if an expert comes in and raises the ID up please raise it up with them.
Hope this helps
If you want to help with an old observation, instead of adding a narrow ID “taxon X” when you are not sure at all, @tag one or a few identifiers that have already identified many “taxon X” observations and just let them identify the observation.
I wouldn’t make a difference between my observations and other people’s observations, for the reason explained by @DianaStuder.
The problem with the CV is that it doesn’t know what it doesn’t know. Often only one or two species from a genus will be included in the model, so it will think all observations of that genus are one of those species. It can also create a cycle of misIDs because people will select the wrong species if the CV is confident it’s right, and those observations go into the next set of training data with the wrong ID, causing the CV to be even more falsely confident on that species in the future. For these reasons, if I identify my observations entirely using the CV, I usually choose the genus of the most confident suggestion rather than the species. Sometimes I even go to the species pages of multiple top suggestions and work up the taxonomy until I find something that matches all of them.
I agree with what others have said, ie do this with your observations but not with others’. One thing to keep in mind is that the other observers can see the CV suggestions too, and there may be a reason they didn’t pick them. I know personally, when I’m uploading out in the field, I take the CV suggestion if it seems reasonable, then check more closely when I get home to see if I agree with it. If I reject the CV suggestions entirely and ID something as “life” or some other broad category, it’s because I know or strongly suspect that the CV is wrong. So if someone went through my observations and added the CV suggestion to everything currently at a broad taxon, they’d essentially be asking the CV to identify everything that I’ve decided the CV can’t identify.
My take is less about whether it’s your IDs or someone elses, but more whether it’s an initial or a confirming ID. I’m comfortable suggesting an ID based on CV but don’t think it’s reasonable to confirm an ID based on CV without at least doing due diligence using the “compare” function. Many naive users fail to ID their own photos of offer excessively (in my view) conservative guesses about species ID. More specific guesses get faster feedback than less-specific guesses. I’d rather have them get a quick guess at rough clade (or appropriately-assertive guess at species) than languish for weeks.
Two further caveats:
First, as the AI has improved, I’ve sometimes tried to refine old broad IDs to species-level based on newer cv suggestions. My hit rate with this strategy has been notably lower: while it’s moved some observations forward, in more cases I’ve had to retract a bad guess. I generally disrecommend this.
Second, in some clades there are active mega-identifiers (e.g. among wasps). I’ve been scolded for prematurely making a species-level guess in these clades and now try to be more conservative. Conversely, in rarely-IDed species like plants a guess can help to quickly rule things out, and has received less pushback.
I view the AI suggestions as a useful starting point to bring possibilities to your attention for taxa you aren’t familiar with, not a way to actually make an identification or choose the answer for you. This is especially true if you are providing IDs for other people. When someone leaves their observation at a broad ID level, it’s not because they need help seeing the CV suggestion, it’s because they didn’t have enough knowledge or experience to verify it and were hoping that someone else would. If all they wanted was the AI suggestion, they would have added it themselves. If all they used was the AI suggestion, then you aren’t actually independently confirming it by agreeing, you are just duplicating the AI suggestion and bypassing the entire purpose of an independent ID (basically, the AI is voting twice). It can be a bit frustrating to do just enough research to know that your observation will probably be tricky to ID and choose to leave it at a high level, only to have someone who has done even less research ID it too precisely with the AI suggestion you chose not to use.
I think when you are actually ready to ID a certain taxon for someone else, you should be confident enough that the AI doesn’t really influence your decision (because even if it disagrees, you know it is wrong, and if it does agree, it’s already the ID you would have reached). Even if you are the first ID on someone else’s observation and not the ID that brings it to research grade, a pretty big fraction of people on iNat will agree with any identification you give them regardless of how much they know, so it’s best to just assume that will happen and act accordingly.
I think the standards can be a bit lower for IDing your own observations, but in that case I mostly worry that someone will blindly “agree” with my ID when I actually do want someone to double-check it for me. Sometimes I’ll ID at a fairly broad level, then write a more specific ID in the comments/description (i.e. “I know it’s definitely in this clade, and I think it’s part of this subclade/species, but I’m not totally sure”). This can be a good middle-ground between leaving something at a really broad ID and going too specific. None of this is meant to be judgmental or critical of others who may do things differently. I’ve definitely been a bit over-enthusiastic with my IDs before and still get some wrong, this is just the way I try to go about it now.
Sometimes when I’m IDing unknowns, I click on the suggestions tab and it says “Mallard”, when the thing is obviously not a duck, or maybe even an animal.
When posting the initial ID for my own observations, I may know it’s a jumping spider (Salticidae), and AI says it’s in Salticinae, so I accept that. Another time I observed a slug crawling up the side of the house. AI said it’s Megapallifera, but I just put Gastropoda, knowing next to nothing about slug taxonomy. The next IDer put it in Philomycidae, the family that Megapallifera is in.
The first consideration is how the CV is trained. From every genus, only taxa that have a minimum number of observations and photos will be included. Widespread taxa with few observations and rare taxa are excluded.
When the app says “We are pretty sure”, it is only meaningful when all taxa in the genus (preferably family) are included in the CV training set. Otherwise it is a guess.
If everyone accepted the CV “suggestions”, there would be no new taxa added to iNat.
Another restriction is that CV can only deal with individual photos where a lot of taxa needs a set of photos for a confident ID.
The guidelines recommend using the Compare function. It is still based on the same CV, so the taxa presented are limited. I used it after I started and got a lot of them wrong.
The best way to identify is to use the relevant keys, field guides and other resources. Even keys are just a shortcut to identify within an area, any feature in taxa treatments/descriptions can be used.
If you know what it is it in the observation, add an ID.
If you don’t know, you can look it up and identify.
Along the way you could find out what needs to be documented for a genus, making your observations better.
It should be about science, getting a correct identification, not necessarily pushing for research grade.
Best practice imo would be to not use CV alone to make any IDs below family level whether your own or others’ observations. And to never use CV alone on others’ observations.
IDing with CV is not contributing anything useful to iNat. One of the problems is CV will often make IDs even when species or genera cannot be distinguished from photos. So using it creates and perpetuates data errors.
If I’m not familiar with the taxon and/or area, I’ll use CV suggestions as a starting point, then choose a parent taxon rather than go to species. For example, see this screen recording for a photo I took in Singapore, a place I’m not familar with, and place that might have a lot of look-alikes as it’s in the tropics.
I tapped on the likely species suggestion to go to its taxon page, then scrolled down to the taxon chart. I’ll then pick a parent taxon that I’m more sure of, then tap on “Select this taxon”.
You can do similar things in the Suggestions tab of the Identify pop-up.
An identification at too broad a level can feel incorrect. Yes, “Melospiza sparrows” may tehnically be a correct ID, but it does not tell me whether the observation is of the ubiqitous Song Sparrow that I have seen many times, or the Swamp Sparrow I’ve been hoping to find. Should my field giude have a check mark for both species or just for the Song Sparrow? One of these two situations is incorrect, but like Schrodinger’s cat, we don’t know which until it goes to the species level.
I’ll definitely avoid using CV alone for other people’s observations based on the feedback. It sounds like there’s a generally accepted feeling that IDing someone else’s observation comes with a higher standard than IDing your own. This feeling is why I asked in the first place. It is sort of funny that there is a distinction here (to many in this thread), but in some ways it makes sense. I would instinctually trust an identifying ID more than an initial ID of someone’s own photo.
However, it sounds like some here think that even for your own observations, you should not use the CV ID if you aren’t very sure. I totally understand this perspective, but I also will say, if this reflects a “standard” or a “community norm” of iNat, then I think the app makes it way, way too appealing to click a species that the AI is “97%” sure is correct. Especially if there’s no easy checkbox to indicate “this is just a guess.”
But it sounds like perhaps this isn’t an actual community norm. Especially because I do think that picking a “bold” initial ID for your own observations, beyond what you can actually identify, gets experts to actually see it more often, where otherwise it might just languish in “insects” forever because that’s often all I’m confident about. Many (most?) of us are just not at the stage in our naturalist journeys to be using many (or any) field guides yet, and I don’t believe iNat has a tool yet to connect us with those keys and resources.
(But I do hear the perspective that this bold IDing of one’s own observations can also get others to blindly “agree” because they trust the initial ID? That agreeing sounds like the bigger problem to me in that case though.)
I know a lot of folks here in the forum are dedicated experts. Hopefully my perspective as a casual hobbyist is a helpful addition. I’m sure there are a lot of folks who are like me who just aren’t bothering to check the forum for what the community rules/norms around IDing are, and are just doing what “feels” most helpful.
I mostly ID a taxon that the CV gets very wrong a significant portion of the time. It is often harder to correct a wrong ID than to refine a broader one. It is also considerably more stressful and discouraging to constantly be seeing bad information and having to disagree with IDs rather than the more positive experience of helping provide new insights. I understand why the CV suggestions are seductive, but from the perspective of an IDer, I would love it if people trusted it a little bit less and instead chose somewhat broader IDs on their own observations when they can’t evaluate the CV suggestions.
I agree that the interface probably makes rather it too easy to just select the top suggestion, but I think you are creating a false dichotomy here. You have lots of options between picking the top species suggestion or IDing as “insects”.
You also don’t have to have a field guide in order to critically assess the CV suggestions and use them to find broader taxa that you recognize. For example, if all of the suggestions look very similar to you, this is probably a sign that you should be cautious selecting an ID. You can also click on each of the suggestions and it will take you to a page where you can see it in a taxonomic tree. A lot of the time you will probably notice that several of the CV suggestions have shared parent taxa – say, if all of the suggestions are for beetles, you are probably safe calling it a beetle. Another option is to look at the top suggestion and go a few steps higher up the tree and then click through to look at observations for your region. If there are multiple other species in addition to the suggested one that look pretty similar to you, it may be wise to select something broader. Similarly, if a large portion of observations are not at species level or not RG, this is often a sign that the taxon is tricky.
As you gain more experience as an observer, you can start to think about the taxonomic relationships of the species you have seen. Do some of them seem more similar than others? What features do they share? How far up the taxonomic tree do you have to go to find shared ancestors? Broad sorting of organisms into groups doesn’t always have to be based on formal learning; humans are good at intuitively grasping resemblances and grouping things into categories. Not all taxa will be intuitive to everyone – likely over time you will discover that there are some that “make sense” to you without much conscious effort and others that you can’t tell apart no matter how many times you see them. This is OK.
From all I understand so far, I’d think it is rather how it is thougt to be used than a “standard”. In my few months, i personally have seen lots of deviation even from contributors with ten thousands of uploads.
Meaning that data quality already today is at risk. I don’t believe that the folks writing and reading here are representing a “standard” user. Those people care about quality, meta communication, consequences to their actions etc. That is not given to everybody.
And yes, i also share your view that it is to easy to consent. And i believe that part of the gen AI buzz is also about the expectation it will strengthen this bad practice. People see an issue already today and have no idea how to prevent it will make it worse.