The bots, when tested alone, performed impressively - correctly identifying medical conditions 95% of the time. But when real ...