Machine learning could fuel a reproducibility crisis in science

"The high accuracy of predictions in tests often fools people into thinking the models are picking up on the 'true structure of the problem' in a human-like way..."

A form of confirmation bias?