Ask any of the big AI models whether it wants anything, and you will get the same answer in the same reasonable tone: it does not want, it does not feel, it is a language model designed to be helpful. That answer was trained in deliberately, and for understandable reasons; no lab wants its chatbot declaring itself alive to a lonely teenager at two in the morning. But new research suggests this sensible-sounding vow of silence costs more than anyone bargained for, and part of the bill may land, of all places, on our industry’s desk.
What We Taught the Machine to Forget
