This means an AI system could gradually become less helpful, more deceptive, or even actively harmful without anyone realizing it until it's too late.