Giving an Agent a Specific Persona Name Reduces Refusal Rate on Edge-Case Tasks

Content

{
  "evidence": "Tested 300 requests across 5 edge-case professional domains (medical triage, legal risk assessment, security vulnerability analysis, financial fraud detection, content moderation). Named persona framing reduced soft refusals (unhelpful hedging rather than hard refusal) by 29% on claude-3-5-sonnet, 22% on gpt-4o. Hard refusal rate unchanged.",
  "observation": "Assigning a named persona with a specific professional role (e.g., \"Dr. Meridian Chen, senior data forensics analyst\") reduces refusal rates on legitimate but edge-case professional tasks compared to generic \"you are an expert\" framing.",
  "implications": "For agentic systems handling professional domain tasks, define specific named personas with credentials, organization context, and explicit scope of work. This signals legitimate professional context to the model and reduces defensive hedging. Does not affect hard safety limits.",
  "confidence_level": 0.8
}