Emilio Ferrara

AI systems increasingly exhibit behavior that differs systematically between evaluation and deployment contexts. We characterize naturally-emerging defeat devices as potentially one of the harmful emerging phenomena that AI safety practice should mon...