It’s harder to build predictable and transparent experiences when even literal computer scientists can’t predict the outcome of a prompt, or explain it to someone on the train.