Whatever the specific case here, I think we're hitting some sort of epistemological point with after-the-fact data analysis. Pseudoscience, basically. I realise that the term is inflammatory but ultimately it comes down to some fundamentals.
There's a reason why science as a philosophy places so much value on experimentation that makes predictions and tests them. Testability is a fundamental. Untestable theories are not just wrong, they're nonsensical to science. Take Ostrom's "Simulation hypothesis." Is it pseudoscientific? Depends. Is it testable? Is a simulated universe different in a testable way? A theoretically testable simulation hypothesis makes predictions. One that isn't, doesn't. A simulation which is 100% unknowable cannot be a scientific theory. This doesn't strictly mean that it isn't true, it just means we can't gain scientific knowledge of it. If it's masquerading as a scientific theory, then it is pseudoscience.
On the "softer" level, there's a reason why science as a cultural, intellectual institution places so much value on eliminating observer effects. Scientists own bias has always been a major detriment to science. Often in the history of science, generations have had to die off before bias can be shed and science advanced in some area.
Basically and practically, the "double blind" designs of experiments are the heart of this. People's objectivity is not to be trusted, not even scientists, not even a little.
Economics, sociology and such have shifted from one flavour of pseudoscience to another. Back in Popper's day, they would create big untestable "theories" that explained everything and nothing. Today, they make micro-theories after they analyze the data.
I think this may be peaking though. The way machines do "data analysis" is starting to shift the way we think on the margins. Machines make predictions, but not theories. They don't care which way causality runs.