“An algorithm that expedites care to a stroke patient in a chaotic emergency room (ER) has a good chance of adoption. An algorithm that reads a routine scan and provides some quantification of what the physicians can already estimate won’t be in as much demand. There are good reasons for algorithms to parse patient records to look for signs of rare diseases, but there are fewer good reasons for using them to evaluate clinical symptoms. It’s cool that AI tools can make diagnoses from scratch, but for most clinical encounters doctors are already pretty good at it.”