Using AI for User Representation: An Analysis of 83 Persona Prompts

This paper analyzes 83 persona prompts from 27 research articles to reveal that while researchers increasingly use LLMs to generate structured, data-driven user personas, these prompts often prioritize brevity and single profiles over the rich, diverse representations traditional to the field.

The Gist

Imagine you are a chef trying to create the perfect recipe for a new dish. In the old days, you would taste-test ingredients, talk to diners, and carefully write down a recipe based on real experiences. This is like traditional User Personas: fictional characters built from real data about real people, used by designers to understand who they are building for.

Now, imagine you have a super-smart, magical sous-chef (an AI or Large Language Model) who can instantly write recipes for you. But here's the catch: you have to tell the sous-chef exactly what to do. If you say, "Write a recipe," you might get a generic, boring dish. If you say, "Write a spicy vegan recipe for a 30-year-old chef who loves jazz," you get something specific.

This paper is a taste-test of the instructions (called "prompts") that researchers are giving to these AI sous-chefs to create user personas. The authors looked at 83 different instructions from 27 recent research papers to see what's working, what's weird, and what might be going wrong.

Here is the breakdown of their findings, served up with some analogies:

1. The "One-Off" Problem

The Finding: Most researchers ask the AI to create just one persona at a time.
The Analogy: Imagine a movie director asking an actor to play one character for a whole movie, rather than casting a whole crew to represent different types of people.
Why it matters: Real life is diverse. If you only build a persona for "John, the 30-year-old tech guy," you might forget "Sarah, the 60-year-old grandmother." The paper warns that asking for just one person creates a very narrow view of the world, missing the full spectrum of users.

2. The "Speed vs. Depth" Dilemma

The Finding: Many researchers tell the AI to be short and concise (e.g., "give me a 3-sentence description").
The Analogy: It's like asking a biographer to write a life story of a famous person, but telling them, "Make it a tweet."
Why it matters: Traditional personas are rich, detailed, and emotional—they help you feel what the user feels. By forcing the AI to be brief, researchers are getting "data summaries" instead of "human stories." They are trading depth for speed, which might make the personas less useful for understanding real human needs.

3. The "Robot vs. Human" Format

The Finding: Most prompts ask the AI to output the persona in structured formats like JSON (a code-like list of data) or tables, rather than a story.
The Analogy: It's like asking a painter to give you a spreadsheet of colors and hex codes instead of a painting.
Why it matters: This turns the persona into a database entry rather than a character. While this is great for computers to read, it might make it harder for human designers to connect emotionally with the user. The paper suggests we are treating people like data points rather than people.

4. The "Demographic Obsession"

The Finding: Almost every single prompt asks for demographics (Age, Name, Job, Gender).
The Analogy: It's like filling out a job application where you list your height and shoe size, but forget to mention your hobbies, fears, or dreams.
Why it matters: While age and job are important, the paper notes that these prompts often miss the "soul" of the user—their attitudes, behaviors, and feelings. The AI is good at guessing "35-year-old male," but it's less good at guessing "anxious parent who loves gardening."

5. The "Black Box" of Instructions

The Finding: Researchers are using very complex, multi-step instructions. Some use up to 12 different prompts in a row to build one persona.
The Analogy: Imagine a Rube Goldberg machine where you push a ball, it hits a lever, which drops a cup, which triggers a fan, which finally opens a door. If one step fails, the door never opens, and you don't know why.
Why it matters: When researchers chain these prompts together, it becomes very hard to tell which instruction caused a specific result. If the AI creates a biased or weird persona, it's hard to know if it was the first instruction or the twelfth that messed it up.

6. The "Magic 8-Ball" Effect

The Finding: Researchers are using these AI personas to predict things (e.g., "How would this user react to this ad?").
The Analogy: It's like asking a crystal ball to predict the future based on a made-up character.
Why it matters: This is a new and risky trend. If the AI persona is based on stereotypes rather than real data, the predictions will be wrong, leading designers to make bad decisions.

The Big Takeaway

The paper concludes that while AI is a powerful tool (a "super-sous-chef"), we are currently using it a bit clumsily.

• We are too focused on speed: We want quick, short answers instead of deep, rich stories.
• We are too focused on data: We want spreadsheets instead of characters.
• We are too focused on one person: We are forgetting the diversity of the real world.

The Advice: If you want to use AI to create user personas, don't just ask for a quick list of facts. Feed the AI real data about real people, ask for rich stories (not just bullet points), and remember to cast a whole cast of characters to represent the diversity of your users. Otherwise, you might end up designing for a robot's idea of a human, rather than a real human.

Technical Summary

1. Problem Statement

The paper addresses the rapid integration of Large Language Models (LLMs) and Generative AI (GenAI) into User Persona development. While personas are traditionally fictitious representations of user segments based on real data to guide design decisions, the shift toward computational generation introduces new risks:

• Lack of Guidelines: There is a scarcity of systematic guidelines or "best practices" for prompting LLMs specifically for persona generation.
• Risk of Harm: Improper prompting can lead to algorithmic bias, stereotyping, loss of human agency, and the generation of personas that lack empirical grounding (i.e., relying on the LLM's internal "knowledge" rather than primary user data).
• Research Gap: There is no existing systematic mapping of how researchers currently utilize prompts to generate personas, nor is there empirical evidence on how these prompt designs affect the quality and characteristics of the resulting personas.

2. Methodology

The authors conducted a quantitative content analysis of persona prompts extracted from academic literature.

• Data Source: A corpus of 52 research articles identified through a Systematic Literature Review (SLR) focusing on GenAI applications in persona development (published within the last 2 years).
• Extraction Process:
  • The team extracted 83 distinct persona prompts from 27 articles (52% of the corpus) that contained explicit prompt text or descriptions.
  • Prompts were treated as single entries even if an article used multiple variations, preserving the holistic approach of the researchers.
• Coding Framework: Two authors collaboratively coded the prompts using a structured framework to answer three Research Questions (RQs):
  • RQ1: Why do researchers use persona prompts? (Purpose/Application)
  • RQ2: How do researchers use persona prompts? (Technical implementation, models, complexity)
  • RQ3: What kind of personas are generated? (Output format, attributes, richness)
• Reliability: Inter-coder reliability was ensured through cross-checking and discussion of disagreements.

3. Key Contributions

• First Systematic Mapping: This is the first study to systematically analyze the landscape of persona prompts in academic literature, providing a baseline for future research.
• Prompt Taxonomy: The study categorizes prompts based on their structure (e.g., single vs. multi-stage), output format (structured vs. unstructured), and content richness.
• Identification of Deviations: It highlights significant deviations from traditional persona creation traditions (e.g., preference for brevity over richness, lack of visual elements).
• Open Dataset: The authors made the collection of 83 prompts publicly available to facilitate further empirical analysis and the development of evidence-based guidelines.

4. Key Results

RQ1: Why are prompts used?

• Primary Use: 81.5% of studies used prompts for persona generation.
• Secondary Uses: 25.9% for prediction (e.g., predicting user responses) and 14.8% for evaluation.
• Applications: Ranged from educational tools and storytelling to bias detection and commercial design.
• Output Structure: 51.85% of prompts required structured output (primarily JSON), indicating a shift toward treating personas as data objects for downstream analysis rather than narrative tools for empathy.

RQ2: How are prompts used?

• Model Dominance: GPT models were used in 76.1% of instances. Cross-model comparison is rare (only 22% of studies used multiple models).
• Complexity:
  • Prompts varied from simple one-liners to complex, multi-stage systems (up to 12 prompts in a single study).
  • Average prompt length was ~107 words.
  • Multi-prompt Strategies: 63% of articles used multiple prompts (e.g., one for skeletal data, one for narrative expansion, one for image generation).
• Data Integration: 74.1% of prompts included dynamically inserted data or variables (e.g., via Retrieval Augmented Generation), moving toward "computational personas."
• Transparency: Only 29.6% of articles disclosed hyperparameter values, suggesting many researchers rely on default settings without optimization.

RQ3: What kind of personas are generated?

• Format:
  • Text (96.3%) and Numbers (66.7%) are the dominant formats.
  • Images are rarely generated (7.41%), despite being a staple of traditional personas.
  • Only 7.41% of entries produced a "rounded" persona containing text, numbers, and images.
• Quantity: Most prompts (70.6%) requested a single persona, deviating from the traditional need to represent diverse user segments.
• Richness & Attributes:
  • LLM-generated personas are less information-rich than traditional data-driven templates (Average 5.48 attributes vs. 8.83 in traditional templates).
  • Demographics are the most requested attribute category (52% of all attributes), followed by Contextual Information (11.3%) and Attitudes (12.7%).
  • Brevity vs. Richness: Many prompts explicitly requested short descriptions (e.g., "3-5 sentences"), conflicting with the goal of creating detailed, rounded user profiles.

5. Significance and Implications

• Shift in Paradigm: The field is moving from "narrative-based" personas to "data-object" personas (JSON/structured), which aids analysis but may reduce the empathetic utility of personas for designers.
• Risk of Stereotyping: The heavy reliance on demographic attributes (age, gender, occupation) without sufficient behavioral or attitudinal depth may reinforce stereotypes rather than capture complex user needs.
• Evaluation Challenges: The use of complex, chained prompt systems makes it difficult for humans to evaluate the causal link between specific prompt instructions and the final persona output, creating "hidden pockets" of error.
• Recommendations:
  • Researchers must ensure primary user data is integrated into prompts to maintain the "data-driven" nature of personas.
  • There is a need to familiarize computational researchers with persona theory (empathy, diversity, representativeness) to avoid "shallow" prompts.
  • Future work should focus on establishing evidence-based guidelines for prompt design and evaluating the impact of different prompting strategies on persona quality.

In conclusion, while LLMs offer efficiency in persona generation, current practices often prioritize speed and structure over the depth, diversity, and narrative richness required for effective user-centered design.