How can we scientifically assess knowledge contained in a large language model?