There is great statistical power in these ML models, in many cases the "random shit" will become the topic of conversation due to shared social factors that can be predicted, you simply neglect to recognize all the times the modeling failed.