Telling an AI model that it's an expert makes it worse

A USC research paper finds that persona-based prompting — telling an LLM it is an expert — actually hurts performance on factual and coding tasks while improving alignment-dependent tasks like safety and writing. On the MMLU benchmark, expert personas reduced accuracy from 71.6% to 68.0%. The explanation is that persona prefixes activate instruction-following mode at the expense of factual recall. The researchers propose PRISM, a gated LoRA mechanism that selectively applies persona-based behavior only where it helps, falling back to the base model for knowledge-dependent tasks. The practical takeaway: for accuracy and facts, skip the persona; for alignment, safety, or structured output, specific persona guidance can help.