[iamstupidsimple]

The input to these algorithms are usually human understandable and quantifiable signals like likes, text sentiment, maybe engagement history -- and the output is probably a score than can be ranked. Ultimately though even if the algorithm is a black box (entirely possible it's not ML based!) we can still evaluate it in a lab environment.

Some of the signals might be generated by ML also, like photo labels, but ultimately these things are very understandable if you have the model and data.