How does AI identify and correct its biases?