I have used TP for a number of years (free and paid), including, for a brief period, as a coach. I didn’t count individual training plans, but it is a lot.
I also have experience in ML. Key is automated generation of datasets, in particular metadata. At my old job, dataset generation was 90+ % of the work. No exaggeration. That was because of a multitude of factors, e. g. that the workflows were not designed with long-term data management and retention in mind. The changes we proposed would have meant more work for the process engineers who are the ones generating the data. Plus, there were lots of other complicating factors (tools from different manufacturers that lack key capabilities, no control over their software, etc.).
Seeing how, hmmm, old school TP is, I wouldn’t count on the necessary tech being in place.
Who is they? TP? How many of those are cycling workouts? How many training plans are comparable? And most importantly, how much effort would it be to generate datasets for scientific analysis (by ML or otherwise)?
No, but dataset generation probably is, which is the lion‘s share of the work. Roughly speaking, if you want to do ML in a systematic fashion, you need
- automated dataset generation based on certain criteria and
- good analysis tools for the entire dataset.
(This is based on my own experience in that area — admittedly in a completely different industry, but I think these paragraphs also apply to TR.)
These tools are totally invisible to the end user. TR has started experimenting with ML in 2015, I think (going from memory here). And the ML-based features it has rolled out indicate to me what kind of capability they have in the dataset generation and analysis workflow. They first rolled out Progression Levels, which indicates to me that at that time, they could select single workouts that match certain criteria in order to answer questions like “Of those two comparable workouts, which is harder and by how much?”
The latest feature is an ML-based Plan Builder that proposes e. g. training volume and number of intense days based on your training history. To me this suggests that TR can now create datasets consisting of entire training histories that match certain criteria. That isn’t easy, you need to specify and implement tons of boring stuff such as
- how to package the data,
- how to keep track, save and standardize metadata,
- write custom analysis software that makes use of these standards,
- have an efficient interface to the database, etc.
All of that takes years of development as there are tons of stakeholders (because you might need to make changes to the database, etc.). And then you can ask the scientifically interesting questions.
Given TP’s comments you posted, I am led to believe that none of that infrastructure is in place whereas TR’s released features suggest you can do that with TR’s data pool. Hence, my comment that TR’s data pool is a gold mine.
But maybe the comparison to a gold mine is not the right one since it takes significant effort to mine and refine gold ore. Maybe TP is the gold mine (or palladium mine), and TR is a complicated water tap by comparison? 
ML can definitely help coaches. How many times have people recreated the same workout in TP? A lot of coaches would likely welcome something like AT if they knew what it would do and offer enough knobs to tweak things.
IMHO their idea to not use ML will eventually put them out of business.