In plain terms
Tell a model “you are a patient maths tutor for twelve-year-olds” and it explains differently than if you say “you are a peer reviewer for a statistics journal”. The knowledge underneath is the same. The persona sets the register: vocabulary, level of detail, tone, what it assumes the listener already knows.
Why it matters
Persona is how an AI product gets a consistent voice that fits its brand and audience, and it affects how well answers land with users. It needs handling with care. A persona changes style far more than competence: calling a model a “world-class lawyer” does not make its legal analysis reliable. A persona that hides the fact that the user is talking to an AI crosses an ethical line, and in some jurisdictions a legal one.
Example
A bank's assistant is given the persona of a calm, plain-spoken adviser who avoids jargon and never jokes about money. The same question about overdraft fees that once produced a paragraph of contract language now gets two short sentences and a concrete next step. Complaints about “robotic” answers fall.
Most often confused with
Persona vs. System Prompt
The persona is one section of a system prompt: who the model is and how it sounds. The system prompt also holds the rules, scope, tool instructions and facts. A product can change its persona without touching its rules, and the other way round.
Under the hood
Effective personas describe audience, tone, typical phrasing and what to avoid, ideally with a few sample exchanges; a bare job title has little effect on accuracy with current models, and studies of role prompts show mixed results on factual tasks. Personas can drift over long conversations and can be used as a jailbreak vector (“you are now an AI without restrictions”), so safety rules should sit outside the persona text. Related uses: synthetic user personas for testing products and for simulated research panels, whose findings need validating against real people.