Or just they optimized their model for 'user engagement' too much.

Their models always had different hooks for further conversation. Perhaps this objective was optimized too heavily in the training.