Large language models combined with confidence scores help them recognize uncertainty. That could be key to making robots safe and trustworthy.