We show that a GPT-3 model can learn to express uncertainty about its own answers in natural language -- without use of model logits.