Telling ChatGPT to behave like a hostile auditor can make its answers more transparent and stop hallucinations