"The research from the Stanford-Berkeley team shows empirically that it isn’t just an anecdotal impression. The chatbot has become empirically worse at certain functions, including calculating math questions, answering medical questions and generating code."