Overview
This systematic review analyzes 52 research articles to map the rapidly evolving landscape of Generative AI (GenAI) in persona development. The study reveals a field currently dominated by OpenAI’s GPT models that is rapidly shifting from static user profiles to "living," interactive agents. However, this transition highlights critical gaps in standardized evaluation and ethical safeguards that must be addressed as the technology matures.
The Problem: Automating the Persona Lifecycle
In Human-Computer Interaction (HCI), personas—fictitious characters representing archetypal users—are essential for user-centered design. Traditionally, creating them is a grueling, manual process involving qualitative data collection and thematic analysis.
While Large Language Models (LLMs) offer a tempting shortcut to automate this, the transition has introduced "wild west" conditions. Researchers are currently using disparate models, prompting strategies, and data sources without a unified framework. This paper addresses the lack of systematic understanding regarding how GenAI is being integrated into the persona lifecycle, the methodologies being employed, and the unique ethical risks—such as algorithmic bias and "hallucinated" user needs—that come with synthetic user representations.
Methodology
The authors conducted a Systematic Literature Review (SLR) following the PRISMA guidelines. They screened 573 initial records from six major academic databases (including ACM, IEEE, and arXiv), ultimately narrowing the corpus down to 52 core articles published between 2022 and 2024.
The research team developed a rigorous data extraction sheet with 35 variables to code each paper across six primary dimensions: basic information, resource availability, research methodology, persona development stages, GenAI/LLM usage patterns, and overall results/discussion.
Key Insights and Trends
The GPT Monoculture
A staggering 82.7% of the studies rely on OpenAI’s GPT family. The authors warn that this creates a "single-vendor dependency," risking the embedding of specific Western-centric biases into the fabric of global user research.
From Static Profiles to "Agentic" Personas
The field is moving away from PDF "trading cards" toward interactive chatbots. Designers can now "interview" a persona in real-time, using LLMs to simulate user responses. A novel trend is representing personas as JSON structures rather than narratives, allowing them to be programmatically injected into other AI workflows or design tools.
The Evaluation Vacuum
Nearly half of the reviewed studies failed to clearly explain how they validated their AI-generated personas. When evaluation did occur, it was often subjective, lacking objective metrics for consistency or "hallucination" detection. Furthermore, while humans are often used for "writing" tasks (40.4%), the actual workflow of how a designer should steer an AI to ensure an empathetic and accurate persona remains underdeveloped.
Detailed Analysis of the Literature
- Publication Trends: The field is exploding; 80.8% of the reviewed papers were published in 2024 alone.
- Data Sources: Studies utilized real user data (53.8%), purely synthetic data (28.8%), and mixed data (17.3%).
- Reproducibility Crisis: Only 26.9% of studies shared their implementation code, and only 28.8% shared their datasets. This lack of transparency makes it difficult to verify if a persona is truly representative or just a "stochastic parrot" of the training data.
- Prompting Strategies: The review identified four distinct strategies: attribute-specific (structured parameters), role-playing (persona perspectives), data-driven (grounded in datasets), and creative (psychological depth).
Limitations and Future Work
The authors acknowledge that because the field moves so fast, this review is a "snapshot" that may soon be outdated. Key limitations in current research include a lack of longitudinal studies measuring if AI personas actually lead to better final products and the "Black Box" problem of closed models obscuring training data.
Future Work Directions: * Model-Agnostic Verification: Comparing outputs across GPT, Claude, and Llama to identify model-specific biases. * RAG-based Personas: Using Retrieval-Augmented Generation to ground personas in specific, private datasets to prevent hallucinations. * Standardized Frameworks: Developing a multidimensional evaluation framework that combines perception metrics, factual verification, and behavioral consistency testing.
Summary updated