Should we be training an AI model that will put us out of work?

In the few years I’ve been on iNat I’ve watched with both pride and worry as Computer Vision suggestions get better and better.

My pride has mostly turned to worry with the rise of LLMs and generative AI, and the number of people losing jobs as a result.

I am starting to wonder if, as a professional taxonomist, I am not doing both my profession and new generations of taxonomists a disservice by adding correct IDs that are used to train AI. At some point iNaturalist is going to end up with the best dataset of imagery with the best identifications that can be applied today. This will be extremely valuable to train AI, and such a trained AI can be used to put taxonomists out of their jobs.

What do other taxonomists think about this?


AI can’t identify undescribed species if taxonomists haven’t discovered them yet …


I guess I would say that I don’t believe that taxonomist = identifier. That’s one part of the job of taxonomists, but there are many other elements as well which would be more difficult to for AI to replace. Also, for some species/taxa, iNat’s CV is a long way from being reliable enough to replace taxonomists (eg for the many taxa not easily IDed from photographs).


I’m not a taxonomist per se but i’ve messed around with the algorithm a lot lately because i like to use it as a form of autocorrect instead of trying to spell out species and i have to say while it does do a good job for making a plant species list or something, you have to know quite a bit about plants to even know what to point the algorithm at as someone without botany expertise isn’t even going to notice a sedge being different from a grass, etc. I’m more worried about taxonomists breaking the ability to identify plants at all with hypersplitting and such, but maybe that will get fixed too when eventually cheap genetic sequencers can be attached to phones. I don’t know.

I don’t see any harm in training the algorithm with my observations and am happy to do so personally


Since taxonomy has been based historically on morphometrics and meristics in many organisms and in recent decades on genetics, I don’t see any device that relies strictly on a photo replacing human taxonomists.


I am a professional plant taxonomist working at a Herbarium and plant ID is a significant proportion of my job. Describing new taxa is a relatively rare event.


Currently the Computer Vision needs correctly identified observations to learn and doesn’t know species with very few or no observations. I frequently need iNat to import missing names of South American katydid species from external sources, or ask a curator in case this doesn’t work.

And species unknown to the Computer Vision it often misidentifies if there are similar species with numerous observations. For example young and unidentifiable Conocephalini nymphs in North America it seems to identify almost always as Conocephalus strictus, although there are often ten or more Conocephalus and Orchelimum species in the area. And the majority or possibly all observations of Neoconocephalus triops in South America in fact correspond to local species of this very diverse genus. The latter is really a problem, since the real southern distribution limit is unknown.

And there are also observations of undescribed species on iNat.


A lot of plant taxonomic work involves very small characters that require a dissecting scope to properly observe. I’m not terribly concerned about AI replacing me.


I don’t think this is a real issue yet. It benefits more than it harms, at least for now.


CV is only as good as the pictures people take, or can take, and there’s many species that require dissection or microscopy to positively ID. While the CV does an excellent job at some taxa, its still going to fail miserably at anything without a lot of observations, or good photos, or cryptic species, etc etc etc… we’re a long way off from it being able to really nail those down.

Honestly there’s some cases where CV may make some things worse; there are certain extremely common taxa that just tend to end up as dump taxa because unless you’ve done research on the genus, you’re not going to know the difference between the really common species and the much less common species that just happens to have a rare tree association and so might just end up going with the common one on the top of the CV suggestion.

How many Laetiporus huroniensis observations have been incorrectly IDed as Laetiporus sulphureus because people don’t realize that the former grows on conifer? We’ll probably never know.


Progress is good. I like digital cameras. AI can help us learn new things.


This question has been on my mind long since and I was glad to find it posed by someone today - it even prompted me to join the Forum for the first time in 3 years.

I wish there were enough taxonomists to do all the taxonomic work that needs to be done in order to understand and preserve the biodiversity that we still have but are quickly losing. I’m not a taxonomist, not even a biologist, and I feel I’m racing against time in an effort to identify biodiversity hotspots in my region before they’re lost (as a consequence of logging, fires, invasive species or climate change, for example). Way more than half of my IDs have not been validated yet, even though I usually take great care to photograph the most relevant features for identification. So… I guess we still have too few taxonomists out there…

AI is excellent at providing answers that are usually right, but so far it is terrible in assessing its own reliability. It’s not “suspicious about its own conclusions”, as I usually tell my Math students to be. So far, only a human can be fully aware of his or her most likely mistakes in order to double check and prevent them. So far, only a human taxonomist is able to say “I’m aware of all possibilities, I have excluded all but one, so that one has to be the right one”. But eventually the algorithms will be accurate enough to do that, of course.

What we have to do in the meantime is to think about fair distribution (the taxonomists who fed the algorithm MUST continue to be paid) and about new tasks for the people whose jobs were [partly] taken. With fair distribution and proper reassignment, it can be a huge opportunity for us to do a lot of pending work in the assessment and protection of biodiversity. The question is: are we collectively oriented to do that work, our leaders included?


Wait, there are jobs for taxonomists!?? :stuck_out_tongue_winking_eye:


I’m not a taxonomist, but of course AI could potentially impact so many jobs - it’s really a crap-shoot no matter your age or profession. These sorts of things should certainly be on our minds, but at worst technological advance is a mixed bag. And there are so many other things to rage against. So for me, since iNaturalist is fun and not obviously evil, I think it’s fine to just live in the moment and enjoy participating.

Even with increasingly knowledgeable AI, I don’t think all hope is lost for professional taxonomists. Look at chess, for example - over the last few decades computers have gotten so they can whup any human. Was it the end of chess? Nope, to the contrary - it’s more popular, and more people make a living at it than ever before. It’s super easy to consult a computer for the best move in any given position, but do people nevertheless still hire coaches? Yep.

Sure, there are many policy reasons for the existence of taxonomists. And AI will cause the profession to change. But ultimately, a lot of the taxonomy arena is built around public interest in nature, and as long as that exists (and AI may increase public interest!) my guess is there won’t be a substantial negative impact on the profession, even if AI causes some of the tasks to change. I guess at some point if a taxonomist’s duties change so much, perhaps a lot of it is no longer taxonomy, right? But I don’t think any shift will be so sudden. And really, is the world better if we can point our phone at a plant and it’ll tell us exactly what it is? Obviously there should be some negotiation on the details, but I say yes.


AI … fifty IDs very good, fiftyfirst ID mistaking an urchin for a spider …


Not a taxonomist … but still waiting for That Someone to formally describe the Table Mountain Strawberry spider we found on a CNC.
Long ‘waiting for description’ queue - which may be discarded when that scientist retires?
Give me a specimen, I might get to it, one day


If I am not mistaken, the company behind the minION nanopore sequencer (which already can be plugged into a laptop and costs „just“ $2000, I think) is currently working on one.

So soon every individual plant can be its own species and we can leave our clunky cameras that always get in the way at home because we‘d just need an entire bio-lab‘s worth of equipment to properly prepare the samples. I‘ve always wanted a field-centrifuge (oh, and don’t forget to bring a cooler for the taq-polymerase and the primers). :D


Computer Vision (CV) is only the start because it requires photos to train. An exciting idea is for the CV to be trained to recognise characters used in dichotomous keys, multi-access keys or just the scientific literature in general. If the ontology can be developed (and I think it can) and the CV can recognise them and “understand” the literature, then suggestions on what a given observation might be could be made even if nobody has ever uploaded a photo based on the scientific descriptions. Words. Would be handy for finding taxa described more than once as well


To answer the question in the OP:

  • Yes, we should be training the AI
  • IMO, AI will not put taxonomists out of work

The discussion in the 17 posts above, reminded me of this taxonomy quote from entomologist Doug Yanega:

"Identifying bugs to species could probably qualify as the slowest biological science.

Case in point: We just got back a loan of 3 bugs that were borrowed in 1958 for a revision.

The original borrower is dead. The person who they went to from there is also dead. The next person who had them just retired.

The curators were unable to find anyone interested in carrying on the revision, so they sent them back to us — still unidentified.

59 years on loan, and no progress made at all.

While I admire the bookkeeping that allowed the specimens to find their way back to us, it’s still sort of sad, and far from unique."


An interesting way to answer this question is to look at the rate of change of observations that Need ID. As long as there is a positive rate of change, there are too few human identifiers out there. In butterflies at least, the rate of change is about +hundreds / day globally.