Grounded personas
A readable guide to the LLM persona-generation research: what a persona can hold, what it quietly flattens, and what grounding in evidence actually buys.
LLM persona generation means telling a language model to answer as a particular kind of person rather than as a generic assistant. Usually the point is to have it stand in for a user during research or testing.
Work published between 2022 and 2026 says two things at once. A model can hold a persona well enough that its answers shift the way a real demographic shifts. It also holds one badly enough that a simulated group agrees with itself far more than real people ever do. Whether a persona is usable comes down to one question: was it built from evidence you can point at, or from adjectives somebody typed?
Here is what the papers actually support, sorted by the four things practitioners keep asking me.
01Can a model hold a persona at all?
Yes, and that part is settled. Salewski and colleagues (NeurIPS 2023) had models impersonate people and then measured what moved. Models asked to be children of different ages reproduced human-like developmental stages of exploration in a multi-armed bandit task. Models impersonating domain experts outscored models given no persona on language-based reasoning. A model prompted to be a man described cars better than one prompted to be a woman, which tells you the steering wheel is wired to the bias as well as to the voice.
Then there’s the part most people skip. Zheng and colleagues (Findings of EMNLP 2024) ran 162 roles across four model families and 2,410 factual questions. Adding a persona to the system prompt did not improve accuracy over no persona at all. Your “you are a senior UX researcher” preamble is changing register and stance. What it isn’t doing is making the model more right.
Park and colleagues answered the architecture side of it in 2023 with generative agents: a memory stream, reflection, and planning, with each piece knocked out in turn to show all three carry the believability. Read that word slowly. Believability was what got measured, so that is what you get.
02Does the persona move outputs the way you want?
Sometimes, and the good result is narrower than people remember. Argyle and colleagues conditioned GPT-3 on socio-demographic backstories drawn from real survey respondents, found the model’s biases to be fine-grained and demographically correlated, and named the property algorithmic fidelity (Political Analysis, 2023). Those are the silicon samples everyone cites.
Bisbee and colleagues ran the same trick and reported the bill (Political Analysis, 2024). They prompted ChatGPT with personas built from real ANES survey respondents. Averages across 11 sociopolitical groups came back close to the human ones.
The trouble was underneath the averages. The simulated people varied far less than real people do. Nearly half the relationships in the data, 48% of regression coefficients, came out statistically different from the real estimates. And the answers moved when the prompt wording changed, or when the model was updated.
So averages survive. The links between one thing and another often do not, and those links are what a study is for.
The most useful number I have read on grounding comes from Park and colleagues (arXiv:2411.10109, v3 June 2026). Working with 1,052 Americans, agents built from participants’ own interviews and surveys reached 86% of those participants’ two-week test-retest consistency. Agents built from demographics alone reached 74%. Same models, same task. Twelve points sat in what the persona was conditioned on.
03Where it breaks
Four named failure modes, each measured by somebody.
Caricature. CoMPosT (EMNLP 2023) breaks this into two faults. Does the model treat everyone in a group as interchangeable? And does it push stereotyped traits past anything you would meet in life? GPT-4 scored badly on both, worst of all for political and marginalised groups.
Skew. Santurkar and colleagues built OpinionsQA across 60 US demographic groups and found model opinions substantially misaligned with what those groups actually think, with people over 65 and widowed respondents among the least represented. Steering the model toward a group did not close the gap.
Flattening. Chen, Zhu and Zheng benchmarked four models against the General Social Survey and the World Values Survey in July 2026. None beat a baseline model at predicting individual responses. All of them treated demographics as far more predictive of attitudes than they are among real people, and on a segment-targeting task they inflated the gaps between segments two to fourfold. Aher and colleagues had already caught the shape in 2023 and called it hyper-accuracy distortion. Qin, Li and Cheng name it from the other side: the average persona, one voice standing in for a whole segment.
Drift, and the urge to agree. Across nine models, identity drift grew with model size, and assigning a persona did not stabilize it (Choi et al., 2024). Abdulhai and colleagues later cut inconsistency by more than 55% using multi-turn reinforcement learning against three consistency metrics, which makes drift an engineering problem rather than a prompting problem. Then stack sycophancy on top. Responses that match a user’s views are more likely to be preferred, by human raters and by preference models, sometimes over correct ones (Sharma et al., 2023). A persona interviewed by the person who wrote it is being asked to please that person.
04What “grounded” actually adds
Salminen, Amin and Jansen read 83 persona prompts from 27 studies. Demographic attributes appeared in nearly all of them, most prompts asked for a single short persona, and 74% pulled in dynamic variables or external data. A demographic label plus adjectives is the field default.
Every result above says that label is the weakest conditioning available to you. Demographics-only agents reached 74% where interview-grounded agents reached 86%. Models overweight demographics by two to four times. Prompting by group label flattens people into caricature.
So “grounded” means something specific here, and it is a rule rather than a mood. Every trait in a persona has to trace to something outside the head of whoever wrote the prompt. Anything that cannot be traced gets cut before the interview runs.
That rule is what I built Grounded Simulation around, written up in Grounded Simulation: Task-Preserving LLM Personas for Synthetic User Research, an SSRN working paper (doi:10.2139/ssrn.6503241) currently under peer review. The architecture has its own plain-language essay, and the workflow piece shows where personas sit in a full study.
Hold three things loosely, because nobody has settled them yet. Whether any conditioning recovers the tails of a population instead of its middle. Whether interview grounding holds up outside the US samples it has been tested on. Whether consistency training buys you fidelity or only smoothness.
05Three moves you can make this week
None of these need the architecture around them.
Replace adjectives with sources. If a trait doesn’t trace to a transcript line, a survey item, or a published instrument, delete it and check whether the persona still says anything worth reading.
Check the spread before you interview anyone. Ten personas that agree with each other on the question under test are one persona wearing ten names. Bisbee’s low-variance finding says they will agree by default, so measure it rather than assuming you dodged it.
Build personas in a session that never sees your hypothesis. Sycophancy plus a leading interviewer is the cheapest known way to measure your own prompt and call it a finding.
What would you need to see inside a persona before you’d let a study built on it change a roadmap? My answer to that, and the current status of every paper behind it, sits on the research page.
06References
- In-Context Impersonation Reveals Large Language Models’ Strengths and Biases
- When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Generative Agents: Interactive Simulacra of Human Behavior
- Out of One, Many: Using Language Models to Simulate Human Samples
- Synthetic Replacements for Human Survey Data? The Perils of Large Language Models
- LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
- CoMPosT: Characterizing and Evaluating Caricature in LLM Simulations
- Whose Opinions Do Language Models Reflect?
- When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses
- Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies
- Examining Identity Drift in Conversations of LLM Agents
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- Towards Understanding Sycophancy in Language Models
- Using AI for User Representation: An Analysis of 83 Persona Prompts
- Restoring Heterogeneity in LLM-based Social Simulation: An Audience Segmentation Approach
- Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?
- Grounded Simulation: Task-Preserving LLM Personas for Synthetic User Research
07Further study
Branching by what you are trying to do next.
If you are writing your first persona prompt.
- Using AI for User Representation — the census of what everyone else is prompting, so you can see the default before you inherit it.
- When “A Helpful Assistant” Is Not Really Helpful — read this before you assume the expert persona is earning its place in your system prompt.
- In-Context Impersonation — what persona steering demonstrably moves, plus the bias it moves alongside.
If you are grounding a persona in real data.
- Out of One, Many — where algorithmic fidelity was defined and tested, and the origin of conditioning on backstories instead of labels.
- LLM Agents Grounded in Self-Reports — the 86% against 74% comparison, and the strongest current case for interviews as grounding.
- Large Language Models as Simulated Economic Agents — homo silicus, for the case that a simulated agent is a modelling instrument rather than a person.
If you are auditing a simulated study somebody else ran.
- CoMPosT — individuation and exaggeration give you two things to check by hand in any transcript.
- Whose Opinions Do Language Models Reflect? — which populations your model was quietly bad at before anybody wrote a persona.
- Examining Identity Drift — ask how long the interviews ran, then ask who checked turn 30 against turn 3.
If you want the strongest case against LLM personas. Bisbee and colleagues: averages match while variance collapses and 48% of coefficients break. Correct means with wrong structure is the failure mode nobody notices in a demo. Chen, Zhu and Zheng: four models, two large surveys, no win over baseline at the individual level, and between-segment gaps inflated two to fourfold. Qin, Li and Cheng name the average persona directly, with partial repairs that trade one dimension of accuracy for another. If those three hold up, the honest scope for personas today is narrower than the market wants it to be, and I would rather you read them before you buy anything, mine included.