Identifying a skink (humans vs AI)
AI. Its definitely a topic in vogue at the moment. Whilst there are many uses of AI in ecology, (e.g. its can help capture, sort and analyse data), some even using AI to draft and write papers. Let me make it clear there are many positive uses of AI in ecology, however I want to limit my scope of this post to the use of AI in species identification. More specifically the use of AI-powered image recognition tools in species identification.
Numerous websites/applications are appearing everywhere, iNaturalist, eBird, BirdNet, Pl@ntNet and I am sure there are more. They have not only come into existence, but now have established userbases, and online communities. It is easy to forget the time of old, where to ID a given species, you relied on keys, field guides, and large taxonomic reference books. But before I bore you with arguments, lets put you in a real-life situation to keep things interesting - then I can write about the ramifications.
Lets take you out on fieldwork! You are on a farm in Western Victoria, on what was once Eucalyptus largiflorens woodland.
You come across a skink in the wild!
Wait do you see it...?
Oh wow, what very nice camouflage! Luckily for you, you have your freely avaliable Key to the skinks of South Australia with you, courtesy of the South Australian Mueseum. I would take the gamble, and hope the key works for West Victorian skinks (given that skinks don't care for state boarders, and I was in closeish distance to SA).
If anyone reading this blog wants to try and ID this skink with a key, here are all the photos you need. Important: do not look at the URL, just copy them onto a word document or equvilant, I realised if you try and download them, the default filename may also give away the answer:
Now let's ID this skink...
Oh it keeps moving!
Key to the Skinks of South Australia, Page 11:
1. a) Each dorsal body and tail scale with a sharp, projecting spine........... Ergenia (page 40)
b) Dorsal scales not spiny but may be keeled or grooved, or completely smooth........... 2
The scales look smooth to me. There are no projecting spines, so it isn't in the Genus Ergenia. Let move on to 2. I will keep a running track of our progress in the key.
1b -> 2
2. a) Head relatively large and distinct from the neck; all limbs short, of similar size with five fingers and toes of roughly equal length; tail obviously shorter than body; moderate to very large size (greater than 12cm length)...........Tiliqua
b) Not as above........... 3
There is definitely no distinct head or neck. Notice that a) describes many morphological features, in general keys are best considered sequentially, in order. Our skink has different length toes, particularly on hind legs, It is also nowhere near the size of a Blue-Tongue lizard! (such as those in the genus Tiliqua)
1b -> 2b -> 3
3. a) Medium to large skinks, adult head and body length mostly more than 50mm; five fingers and toes; lower eyelid scaly and moveable...........4
b) Medium to small, sometimes very slender skinks, adult head and body less than 60mm; some fewer than 5 fingers or toes; lower eyelid with a transparent disc ('Spectacle') either moveable or permanently closed........... 10
Our skink is medium to small, with a head less than 60mm. This is probably enough to progress, but I can also see the diagram on page 10, that shows the different kinds of eyelids possible. Our skink has a Spectacle eyelid.
1b -> 2b -> 3b -> 10
10. a) Skull roof with a single large diamond-shaped scale (interparietal and frontoparietal head scales fused); five fingers and toes; lower eyelid forms an imovable 'spectcle' over the eye...........11
b) Skull roof with either two or three large plates (interparietal distinct from frontoparietals, which may be either paired or fused); limbs may have fewer than five fingers or toes; lower eyelid movable with a disk, or an immovable 'spectacle;............ 12
Now this one is getting a little tricky. How would I know what interparietal scales are, and if they are fused or not? One of the major benefits of a good key, is that it can show you in images/diagrams/drawings. To the credit of the key, this is well illustrated.
In my case, I've zoomed in, rotated and marked up the image previous. Our skink has parietal scales (outlined in blue) in contact with each other, behind a fused interparietal and frontoparietal scale (outlined in yellow). In other words - the skull roof has a single diamond-shaped scale. Now onto 11.
1b -> 2b -> 3b -> 10a -> 11
Key to the Skinks of South Australia, Page 14:
11. a) Head and body flattened; often climbing on vertical tree trunks, rock faces and fences; two or three upper eyelid scales enlarged to form a 'visor' above the eye.............Cryptoblepharus
b) Head and body not flattened; mainly on the ground or fallen wood and litter; upper eyelid scales not projecting, hidden under eyebrow (supraciliary) scales............. Morethia
Well the head and body is fairly flat. But interestingly the key has now picked up on some behavioural traits - we must be close! Our skink is arboreal, basking in the hollow of a large tree trunk. There is a visor above the eye too. It matches the genus description of Cryptoblepharus!
1b -> 2b -> 3b -> 10a -> 11a -> Cryptoblepharus
Now we suddenly move to Page 22, entitled 'Key to the South Australian species of Cryptoblepharus'. We also go back to 1 again.
1. a) Dorsal pattern includes only alternating stripes, with no dorsal light and dark speckling; scales on the soles of the feet dark pigmented, smooth, and rounded.............C. pulcher
b) Dorsal pattern includes light to heavy pattern of light and dark speckling; soles of the feet whitish, with the scales at the base of the toes triangular, projecting and sharply pointed........... 2\
The dorsal scales on the skink have a heavy pattern of light and dark, its speckled too, not stripes. Harder to see, but if you zoom in on the last image, you can just make out the light-coloured scales on the soles of the feet. So we move onto 2.
1b -> 2b -> 3b -> 10a -> 11a -> Cryptoblepharus -> 1b -> 2
2. a) Most often six supraciliaries; subdigital lamellae on 4th toe smooth-edged when viewed laterally.............C. australis
b) Most often five supraciliaries; majority have subdigital scales that are keeled (sharp pointed apex) when viewed laterally, but some have smooth-edged subdigital scales........... 3
I count 5 supraciliaries
1b -> 2b -> 3b -> 10a -> 11a -> Cryptoblepharus -> 1b -> 2b -> 3
3. a) Usually dusty grey to light brown with weak dorsal speckling; and distinct, moderately straight-edged pale dorsolateral stripes; north of the Barrier Highway .............C. ochrus
b) Usually strongly speckled blackish and light grey, giving a chequerboard effect; light dorsolateral stripes ragged-edged due to darker scales; south of the Barrier Highway........... C. pannosus
Our skink is strongly speckled, with light dorsolateral stripes. Our skink is Cryptoblepharus pannosus also known as the Ragged Snake-eyed Skink. You have successfully identified the skink.
1b -> 2b -> 3b -> 10a -> 11a -> Cryptoblepharus -> 1b -> 2b -> 3b -> Cryptoblepharus pannosus
Now let's test the AI
iNaturalist
Well that was quick. With 4 high quality images, iNaturalist identified the genus with a 'pretty sure' level of confidence and within a couple seconds. The top suggestion also matches our own key-based identification. I did give iNaturalist some location data too. iNaturalist is the only website I know that identifies skinks.
Of course there are bias, two main ones are:
- Quality of input - AI is only accurate subject to the quality of the photo. iNaturalist is still photo-based. Other tools, such as eBird and BirdNet allows you to identify birds from recording bird calls. This will likely only improve with time.
- Database bias - AI can only be as good as the data it was trained on. I assume it can only identify species that are already existing in the database (and there is a mechanism to prevent AI creating new species). More advance models may even have population data, factoring in the rarity of a given species.
Honestly, It's astounding how easy and convenient it is to use AI-powered image recognition tools.
Conclusions
Human powered identification (with the assistance of a key)
Okay, okay, I am not going to pretend like that wasn't a process. There were 10 steps and it required thought and effort. However the more times you repeat this process, the faster and easier it becomes to use the key. With enough practice, you eventually become confident 'enough' to prehaps not need the key to ID. Especially for the highly distinctive species, or for those species you have previously come across. This level of 'enough' however is crucial. There are no defined thresholds. There are two major benefits to using keys:
- You can trust the information. Keys are written by experts in the field. Generally people who have studied these species for many years.
- Transparency, if you made a mistake, or unsure, you can always go back and forth in the key, and figure out why the path leads to a particular species. Unlike AI, there is no black-box.
Once you can identify and describe diagnostic morphological features specific to a given species (technically called autapomorhies), and are also able to justify why it cannot be any other species (more on this later), this is when I believe you can be truely confident of the ID. In our example skink, the light and dark speckled colouring of the scales on the dorsal side immediately stands out. But it is also every observation you made, compounded with every other observation, that eliminates doubt it can be any other species. It's the diamond-shaped scale on its head, the specticle eye, the visor on top of the eye, and the 5 supraciliaries, its all these observations combined, that are observed on the one individual skink that makes you certain of its ID. It must be noted these are all visual observations. Humans are not only limited to visual data.
Any good ornithologist can identify bird species from just listening to their calls. Similarly, a botanist can crush the leaves of a peppermint gum (Eucalyptus radiata) and smell the reason why it is called a peppermint gum. I want to highlight how valuable utilising multiple senses is. I fear this will become lost in the AI age and is an underrated skill for any ecologist. The ease in which the human brain can combine these multiple sources of data to then recall a specific species, based on this evidence, is impressive (and doesn’t rely on an internet connection/subscription/electricity).
The ability to identify any given species is predicated on the knowledge of why it cannot be any other species. Essentially, the Karl Popper Philospophy but for species identification. Karl Popper basically said any scientific theory can and should be refuted. He argues it is not possible to prove a theory true, you can only disprove them. Disproof is what really matters.
In a similar vein, Keys are able to eliminate all other possible species, until only one species remains (i.e the testing of your theory) and allows you to potentially disprove your own initial ID. Misidentification can occur when you become complacent, and are overly confident in your own ID. You may even be able to justify your identification by choosing and selecting morphological features that match with your current understanding, whilst intentionally or unintentionally ignoring others (i.e. you are trying to prove that your theory is 'true' and not disprove your 'theory'). Even if we are at a 'pretty sure' level of confidence (as AI would say), there is no harm in double-checking with a key. The process, by design, makes you observe those diagnostic morphological features and able to justify why it cannot be any other species. Without keys humans select from the database of species in your head, the ones you remember. Whereas keys filter from a wide base of all possible species in a certain taxonomic group.
AI powered identification
Provides mostly accurate (subject to bias as per above), instant identification of many species. A main benefit of AI is its accessibility in the modern world. You don’t need to have experts around you, or go on guided walks. It enables citizen science to occur at a much more rapid pace, and at much greater scale. It can also facilitate others to become interested in the natural world around them. Apps/Websites can foster this idea further by gamifying the process. There are now leaderboards to rank who can observe the most species, challenges, competitions and whatnot. An interesting case of extreme birdwatching, where birders compete in a so called 'big year', is well documented here.
The downside is that the end result comes out of a black-box. To give an analogy, I guess its a bit like being in a foreign country where you do not speak the language. However instead of learning the language, you continue to use translator app, repeatedly. You are not in a process of learning and discovery. You may know the name of the species, yet not much else about it. As shown in the example above, by using keys you learn a lot about the anatomy of skinks, but there are other ancillary benefits.
If you are using AI as a first resource, do even have the ability to judge if its right or wrong? Imagine you have never observed Species X, to the extent where you haven't even the faintest idea about it, never even researched it. You are put in a position where you either agree with the ID the AI suggests, or you are back to square one, with no ID. Yes you can run web searches and then research retrospectively - but I would hate to be in a dystopian future where people/ecologists preference the use of AI-powered identification simply because its quick, easy and fast. They haven't learned how to ID species in the world around them. AI may also be used solely out of laziness, lack of resources, or in the worse case, simply because 'thats how its done' and this is the only way you know how to ID.
I only argue this as humans are creatures of habit. Once AI use becomes habitual and part of the process in identification, even if it is used secondarily, just to re-enforce your own views, it is precisely at that point, you are mixing AI intelligence with human intelligence. Suddenly the two are very hard to separate. This isn’t unique to species identification either. It applies when spotify recommends you a new song that you like. Suddenly your very own music preferences are being influenced by AI (and the sense of 'self', your own musical preferences, starts eroding away - something I also very much detest but that is for another post).
I could write more on this subject, however the post is already long enough. Field guides (which I have not even mentioned yet) can provide more rapid identification - however this is for another post.
Overall, it may be harder, slower but I find there is real value in still using keys. Using your brain is good for you, and keys make you appreciate the fine details of life.