As members of the public increasingly turn to AI chatbots to understand their world, even subtle latent biases in the underlying models could affect public understanding of the present—and past.