Oooooh, this by @Shamar is also good:

There is no "learning" in "artificial neural network"

[…]
Such kind of virtual machines are composed of tiny devices that are improperly named "artificial neurons" or "perceptrons", but they are simply vector reducers […]
Such vector reducers can be easily composed in a variety of topologies and programmed in a variety of ways, so that a whole "artificial neural network" can be statistically programmed ("trained", in AI/ML parlance) to approximate one of the possible translations from a vector space to a different one.
Yet, there is no "learning" ongoing: just the iterative tuning of parametric vectors to approximate a certain output. […]
[…] In other words, an "artificial neural network" does not learn anything and it's not a network of neurons or anything like that. It does not understand anything about the data and the output vectors has no inherent meaning: its semantic is always attributed by humans according to their insights about the statistical program they uploaded into the vector mapping machine.
[…]
That's why it's statistical programming: you start with a source dataset (improperly named "training set" in the AI/ML parlance) and, after a compilation process that is specifically designed for that specific virtual machine, you get a binary that such machine can run.
While such binary is not constituted by a sequence of instructions, it's still just an algorithmic transformation of the original source that, despite being expressed as cryptic matrices, still contains all the relevant features of the sources.
The software run by a vector mapping machine is not a "model", since it does not give any insight about the relations it absorbed during its statistical programming. Instead it is just a software like any other that describes a rigorous (if unknown) process that a specific machine has to follow to automatically and deterministically compute (or rather approximate) a desired output from a given input.

The quotes are from the essay published by the author under this licence (which is nōn-FOSS, but it’s still worth a read).