The Arduino has 2 KB of RAM, you can't even record much sound on it, much less do ML.
So I agree, the title is sort of odd, but you absolutely can use Arduino for ML. Personally I think it is clunky and difficult to use compared to modern offerings like PlatformIO or MicroPython, but it's familiar to a lot of people and there is value in that.
You're right, it refers to an ide. That definition is restrictive, it would imply that Arduino is not capable of running anything itself (you need a board for that). Or that I can run anything provided you can compile it from arduino ide(you can run gcc).
I'm sorry but whatever the official description is, for many people including me, runs on an arduino means runs on an at mega with a bootloader and maybe on some kind of specific board with funny connectors.
If you mean runs on a cortexM4 with 8megs of flash, say it I'm sure many people will understand. And doing AI (inference) on such boards has been done and demonstrated a few years ago and is still cool.
You know what ? I could play super Mario Wii u on a cortex m4 .. of course it was tongue in cheek as I was playing a video of the game on a cortex m4 plugged into a VGA screen. But technically I was right, my point being you can be right, and misleading.
Why? The Arduino stack will work on ARM chips - better, cheaper chips than the 328p. Arduino.cc make, sell and officially support ARM boards; there are a huge range of cheap Arduino-compatible boards with modern microcontrollers. The cheapest M4 chips are priced competitively with the 328p and M0 chips are available for as little as $0.24 in quantity.
The 328 is an ancient, inefficient, overpriced chip and I see absolutely no reason to endure it any longer.
Because, despite your obviously correct arguments, the brand name Arduino is pretty diluted and commonly just associated with the range of chipsets that made them popular these days (maybe with the addition of the ESP32/8266s in the last few years). That might change in the future but the common usage of the brand name is pretty independent of their product lines.
Despite its low power its simple enough to wrap your mind around, making it good for intro to ECE and programming.
I'd argue it's more the "ecosystem" today, rather than just the organization/company (or the "board").
In many cases, if you want your microcontroller to gain wide acceptance in the hobbyist community (if you're targeting it), you have to provide a "core" or something similar for the official IDE.
At one time, "arduino" meant "ATMega168/328 carrier board loved by hobbyists and despised by everyone else" - but those days are long, long past.
In fact - I think there's a core available for the SoC on the Raspberry Pi - ah (well, something):
Note that they don't say THE Arduino but AN Arduino. What you're thinking about is probably the Arduino Uno which is the original model.
These days Arduino is a platform made up by a common software (IDE, libraries etc) and a family of products that employ different microcontrollers (Different AtMegas but also Cortex processors), different form size (from the original "big" size to the small Arduino Nano) different amount of memories and different extras.
For example Arduino MKR1000 is AN Arduino employing a 32bit SAMD21 Cortex-M0+
That is Arduino "Uno", not Arduino itself. Arduino now refers a platform and an ecosystem for quick microcontroller prototyping. Almost every popular microcontrollers has arduino API support (either officially or by community).
So the term "Arduino" sounds as "microcontrollers" to me.
It could be enough to perhaps implement a very rudimentary voice recognition, which could indeed be interesting for these low cost devices. And if it is just about being activated by a specific key word.
Speaking of machine learning is perhaps a bit much, but isn't it the same in other applications?
I'm a big fan of everything Lady Adafruit does. She has been and continues to be a huge contributor to the IoT ecosystem and OSS ever since she started her little shop 15 years ago.
I wonder whether it can beat professionals at chess. Last time I checked (a few years ago), software running on ordinary smartphones reached really high ELO. The improvements since Deep Blue are really impressive.
You'd probably need to roll your own lightweight face vectorizer though.