In a paper, OpenAI identifies confident errors in large language models as intentional technical weaknesses. Fixing them requires a rethink within the industry.