Oxford researchers tested GPT-4o in medical scenarios and found a 60-point gap between lab performance (95%) and real-world results (34%). The AI provides correct information...