Can statistical models choose themselves? I am trying to make a criticism of the statement

Of course the statistical models used for prediction don’t choose themselves. Someone has to imagine what factors might be relevant, and there is a great deal of expertise and work that goes into designing and calibrating a statistical model.
(my emphasis)

found in the article The Curious Journalist's Guide to Data.
Of course, it is the scientist who chooses what variables to consider in the model and one might fail to consider a variable that would explain most of the variance in the variable we are trying to explain.
I get confused on the sense of this statement typically when considering methodology of model selection. There are objective(-ish) ways to select models (e.g. minimizing AIC via a backward regression procedure). I also get confused about this statement when thinking of bayesian methods and when thinking about machine learning (e.g. decision trees, classification and other). I suppose part of my inability to make sense of the statement might come from the inability to fully grasp the definition of the term statistical model
Is there truth and is there inaccuracy into the above statement?
 A: 
Is there truth and is there inaccuracy into the above statement?

Regarding the truth, think of the classical statistical modelling approach where a statistician sits down together with a subject-matter expert and they design a model for the phenomenon at hand. A simple example could be a multiple regression model for, say, house prices (something encountered early on in Woodldridge's introductory econometrics textbook). The citation makes a lot of sense for this kind of situation.
Regarding inaccuracy, you could say that automated feature selection in regression, e.g. using information criteria, contradicts the statement. But still the general pool of models from which an automated procedure selects one is designed by the statistician.
With regards to machine learning methods, perhaps they do not quite fit in the classical understanding of statistical models -- but this may be subjective.
A: I agree that the term "model" is too vague and could refer to different degree of precision when describing a ML implementation. I think that there are at least three levels of descriptive precision of the term "model":


*

*top level - for classification that could be K-Means, SVM, NN, etc.

*middle level - for SVM that will refer to the different types of kernels; for NNs, that could be some combination of: convoluted, recursive, SLTM, etc.

*low level - will be the number and size of hidden layers and their specifics - sigmoid, ReLU, pooling, dropout. 


Further, there is the choice of gradient descent algorithm and meta parameters. Maybe some of these choices could be considered a part of the training process, rather than the choice of the model, but it's hard to define where one ends and the other begins.
Even if we only consider the "top level" as "choice of the model", in the end, it still takes expertise and experimentation to get good results. Especially if ML is being applied to a new domain.
