Does telling an AI who it's talking to work better than telling it who it is?
Does framing a persona as audience instead of speaker change results?
This explores whether a prompt that tells the model who it's talking to ("You are talking to a lawyer") works differently from one that tells the model who it is ("You are a lawyer"), and what the wider corpus says about why that might matter.
This explores whether telling a model who its audience is works differently from telling it who it is. The most direct evidence says yes, but only a little. Across 2,410 factual questions and nine models, prompts that named the listener ("You are talking to a lawyer") raised accuracy more than prompts that assigned the speaker's identity ("You are a lawyer") Do audience-framed prompts work better than speaker-framed prompts?. The effect is statistically real but small. Don't treat it as a hack that unlocks hidden expertise. Treat it as a hint about how persona prompts actually work.
That hint gets clearer when you look at what speaker personas fail to do. Telling a model it is someone else seems to change how its output sounds more than what it knows or believes. Persona prompts make models follow trait instructions, yet the underlying gaps in sentiment between groups stay where they were Can persona prompts actually reduce bias in language models?. Conditioning models on detailed personal profiles from more than 200,000 people did not improve predictions about those specific people Does conditioning LLMs on personal profiles improve prediction?. One explanation comes from work on how personas are installed. The model's stable "self" comes from post-training, which ties it to a dominant Assistant identity How stable is the trained Assistant personality in language models?. That trained character is sticky in a way that prompt-level role-play is not Are RLHF personas performed characters or realized dispositions?. Read this way, "You are a lawyer" asks the model to put on a costume over a character that is already set. "You are talking to a lawyer" leaves the trained Assistant alone and only changes who it's speaking to, which may be a more natural adjustment.
The headline-ranking study is a surprising case of the same split. A ten-persona demographic panel ranked Upworthy headline variants worse against real clicks than a plain baseline did Do demographic personas help models rank headlines better?. The baseline asked how likely a typical reader was to click. That is a question about the audience. The study wasn't designed to compare audience and speaker framing, but the pattern fits: imagining the reader worked better than having the model play a set of readers.
The idea of modeling the listener also shows up somewhere quite different. In dialogue research, giving an agent an "imaginary listener" reduced contradictions in its persona without extra training. The agent checks whether each line it's about to say would make its own persona recognizable to someone listening Can imaginary listeners reduce dialogue agent contradictions?. So thinking about the audience doesn't just help with accuracy. It can also help a model stay itself.
The honest limit: the corpus has only one study that tests audience against speaker framing head-to-head, and its effect is small. The broader pattern is better supported. Prompt-assigned identities change surface behavior more than substance, and several separate lines of work find that thinking about the listener does better than having the model play a part.
Sources 7 notes
Testing 2,410 factual questions across nine models, addressing the audience ("You are talking to a lawyer") raised accuracy more than assigning speaker identity ("You are a lawyer"). The effect was statistically significant but small.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Across 208,021 participants in the Psych-201 dataset, conditioning LLMs on participant profiles did not meaningfully improve predictions for specific individuals. The standard technique for individuation produces no measurable gains in person-level forecasting.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
Post-training installs stable dispositional profiles that persist under adversarial pressure, marking them as realized rather than performed. The stickiness of trained personas across conversations distinguishes them from prompt-induced role-play that collapses under jailbreaks.
Show all 7 sources
A zero-shot baseline asking how likely a typical reader is to click ranked Upworthy headline variants significantly better than a ten-persona demographic panel. Persona conditioning appears to introduce systematic bias rather than capture audience-specific insight.
Endowing dialogue agents with an imaginary listener via Rational Speech Acts reduces persona contradiction at inference time without NLI labels or extra training. The agent simulates whether utterances would distinguish its persona from a distractor, suppressing generic or contradictory responses.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models
- When Persona Attributes Improve Population Alignment in Large Language Models
- Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation
- The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
- Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference
- Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations