'Ey up AI

I have frequently mentioned AI, and sit on the fence when it comes to its use. In my opinion it is wrong that people should lose jobs and large businesses should hide behind AI to ignore people's complaints, but that is only my opinion and nothing will change. But when it comes to its use in far more important matters, such as bird identification, I have evidence that it is frequently inaccurate. 

I think most people realise it's not totally accurate, but the problem is relying on it. I so often see posts about identification which say "Merlin says" or "Obsidentify makes it a". The issue is not whether you have the correct identification or not, but whether you're prepared to learn by its mistakes. If you don't want to improve in what you do, fine, but most of us would surely rather learn birds as we go through life than just accept a "guess" from some piece of technology. I have come to realise after over fifty years of having an interest in wildlife that a lot of what we learn is subconscious, which makes it far more valuable than school learning. 

I recently went to see and photograph an Isabelline Wheatear at Winterton, and as I was processing and filing the photos I thought I would do a little non-scientific experiment. I had already used AI during my processing of the photos, and use Obsidentify for moths on a regular basis. I have Merlin, but will only use it for calls, as I have always had difficulty memorising birdsong and my hearing isn't that good. It is often said that if a piece of AI gives a result of 95% or more, you can rely on it. So read on.....

I used Obsidentify on the images on my computer screen, which were quite good, although I say so myself. The first I tried expecting the wrong answer, but I was wrong, it came up Isabelline 85%, Northern 15%. To be fair that is better than many birders. But then as I processed the second shot, I noticed that two white feathers, about to be lost, were largely covering the dark alula, which is a good identification feature of Isabelline Wheatear. I'd forgotten I knew that, but it came back to me whilst looking at the photos. The AI obviously "knew" it, but wasn't able to cope with the white feathers, so it came back with 54% Northern 46% Issy. The next photo (not shown), which didn't show the alula at all, came back 97% Northern, although it was showing a small part of the tail, the tail pattern being another diagnostic feature. To test this, I selected a photo which had the alula masked and the tail side on, whereby the pattern could not be determined. At last, a 100% identification of Northern Wheatear!  Except it was an Isabelline Wheatear, found by a birder using human experience and knowledge, and verified by at least one other (me) who has seen Isabelline Wheatears in several countries, although it was only my third in the UK.

I still find AI useful, and it will improve, but I look on it as an extra field guide, not an answer. If you think back to early field guides, and some these days for more obscure regions of the world, they were almost worse than nothing. Apart from poor plates they rarely showed more than one or two plumages, and for something like an immature gull that is often useless. But now there are specialist guides with much more information. These are certainly better than any app currently available, but most people don't have a library of bird books. I still think the best way to learn is by observation, once the initial ID is settled watch the bird, and when you see another, you may be surprised to find that you can identify it for yourself. Even if you rarely see rarities, at least if you come across something new you will know it's not something you've seen before. If you carry a camera, so much the better. Use AI, but just keep the word "artificial" in mind. Having been a birder for at least fifty years, I have been mothing for about 14. When I started it involved hours of looking through guides, and I still couldn't identify some. But using Obsidentify has made the task much easier, and it is possibly more accurate than Merlin is for birds. My "go-to" check is the Norfolk (or Suffolk) moths website, where there are numerous photos of each species. Some moths show such variation that no field guide is up to the task, but AI is my starting point, and I no longer need anything except my previous experience to identify 85% of what I trap.

There is much new technology since I began birding, most of it excellent, such as thermal imagers and photo processing programs, but total reliance on it results in a far less rewarding experience in my view. I "returned" to birding when I was thirty, (from 12-29 I was doing many other things) and for four years refused to twitch. Then I went a bit beserk, but at least I could identify my quarry for myself because I had a good working knowledge of all the regular species. I've had film cameras, but didn't buy digital until it gave better results, didn't even use autofocus until it got faster or more accurate than me. I'm still learning a lot about birds, and it's a great way of life. Now bits of me are failing (deaf, permanently stiff and cataracts) I'm pleased I can use modern technology, but what happens if technology fails? 

As this is pretty boring, especially to birders with any experience, I've put in some photos taken without modern technology. To a photographer they're not very good, but to a birder with a camera, which most of us are, they're perfectly acceptable.