The article identifies fault lines in the global promise of artificial intelligence (AI) in medicine, where laboratory successes fail in real-world clinical practice. The main problem is contextual errors, when AI models do not apply recommendations to a specific context, such as a medical specialty, geographic location, or socioeconomic factors. These errors arise because the training datasets do not contain sufficient information about clinical decisions. An example is the Watson system, which achieved only 33% agreement with local oncologists in Denmark and did not show improvement in patient outcomes. Google Health's AI for detecting diabetic retinopathy dropped from 90% accuracy in the lab to 20% failure in Thai clinics. The solution is the development of models trained on several specialties, capable of switching contexts in real time. The authors emphasize the need for transparent recommendations that acknowledged uncertainty in order to increase confidence among patients, clinicians, and regulators.