The model had been trained on text from a decade in which certain outcomes were considered inevitable. Its predictions were accurate and confident. The outcomes it had been trained on had not occurred. No one had retrained it. It was still predicting.