Warnings from OpenAI and Anthropic reinforce the need for enforceable safeguards, human oversight and a way to challenge decisions that cause harm.