I have tried a few different plant apps to diagnose issues and identify species. The tech is getting better, but it is far from reliable. Just as chat AI tools provide different results based how questions are asked., I have found that lighting, angles, and background are just as important for photo AI. These pics are all of the same palm. 

I don't understand anything about artificial intelligence, but my son, who has 3 degrees and is a mathematician and physicist, teaches mathematics in the morning and works on the development of artificial intelligence via PC in the afternoon and he always told me that they have overdone everything, while the human brain can create and do other things, the PC can process and do calculations only more quickly.

I have tried to identify palms before and thought of using AI or some reverse image search but it didn't really work too well yet. Most of the time it only tells you the genus and is inaccurate. I feel like manually IDing is better for now

 

Also that is certainly not a date palm or any Phoenix, it doesn't even have spikes...😅

I have been using one form of AI or another since 1992 in grad school.  Its been around a long time, few new algorithms have been developed the last decade.  What has changed is the ability of computers to rapidly handle massive amounts of data needed for neural networks, Machine learning,  and statistical pattern classification methods.  Each algorithm each has its strengths and weaknesses. 

First,  complex AI problems require a lot of data and it should not be too redundant.  For example highly similar pics of an adult palmetto adds little in multiples and they may not help much identifying young palmetto seedlings.   So collecting the data needed is not a simple process, you must account for "within class varations" for each class to represent all the possibilities during training of the AI.  Collecting the data needed is non trivial as you pointed out.  Close in pics, profile pics, lighting variations,  view angle, age of the palm etc has to be accounted for in the data.

second, data in-> # classes out is the over all process.  All the images go in and the classes(species) come out  The simplest scenario is one output class, with membership of yes/no as the output.  In this application you have virtually thousands of output classes(species) which makes this a difficult problem.  

third, some of these unknown species images may not be unique enough to identify.  For example, you will not be able to identify many sabals until they flower so a pre flowering plant is indistinguishable among other sabal species, they look too much alike.  AI cannot solve data which is not unique to the class, no algorithm can and no expert can.

Fourth such high dimensional data require vast independent sets of data not used to train the AI in order to validate the algorithm that it has not just fit the training data.  Complex algorithms can fit training data and still fail on independent predictive data, its very common.    There is a famous case study where the military was trying to identify the presence of tanks in the jungle.  One day they took lots of images of jungle containing tanks and the next day there were many images of jungle with no tanks under the canopy.  The classifier was perfect in training(fitting) the data but failed to reliably indentify images with tanks.  Problem is it was sunny on the day the tanks were present and cloudly on the day they were absent.  Validation showed all the algorithm did was identify a sunny day and distinguish it from a cloudy day.

If you are still reading this post, the final problem is that the financial motivation for solving complex problems must be present to justify pursuit of the palm species classifier.  Even if it is all possible to solve, absence of funding will prevent it from happening.  I can think of no significant financial motivation for creating an expensive AI classifier when there is no substantial monetization.