Not at Facebook, but I used to work on an ML system that took well-defined and free-form JSON data and ran ML on it. Both were used in training and classification. Unless a human looked, we had no idea what those custom fields were. We also had customers lie about what the fields represent for valid and less valid reasons.

Without knowing how it works at Facebook, it's quite possible the data points got slurped in, the models found meaning in the data and acted on it, and no human knew anything about it.

How it happened internally is irrelevant to whether Facebook is responsible. Deploying systems they do not properly control or understand does not shield against legal or normal responsibilities!

There is a trail of people who signed off on this implementation. It is the fault of one or more people, not machines.