When Riley Goodside starts talking with the artificial-intelligence system GPT-3, he likes to first establish his dominance. When Jessica Rumbelow and Matthew Watkins, researchers with the machine-learning group SERI-MATS, tried to prompt AI systems to explain how they represented concepts such as “girl” or “science,” they discovered that a small set of obscure terms, such as “SolidGoldMagikarp,” tended to induce what they called a “mysterious failure mode” — most notably, a garbled stream of profane insults. “It’s a mode of communicating in the meeting place for the human and machine mind,” he said