Synthetic personas are artificial entities generated by large language models (LLMs) to simulate human responses. Studies have found that they can mimic human decision-making with a moderate to high degree of accuracy, and they offer a tool for exploring public responses when traditional approaches are constrained by time, cost, or feasibility. Now, Strategy&’s Ideation Center has extended this research for Middle East policymakers by studying the ability of synthetic personas to predict policy sentiment in three Gulf countries.
Key findings:
AI predicted public policy sentiment with over 90% directional accuracy.
GPT-4o showed the closest match to human responses.
The findings held across 60 policies and three GCC countries.
Simple personas were nearly as effective as detailed profiles.
AI captured overall sentiment more effectively than the diversity of public opinion.
Synthetic personas can accelerate evidence-based policymaking.
Synthetic personas are well suited for early-stage policy testing and screening.
Human engagement remains essential where understanding differences across population groups is required.
Our study found that agreement between human and synthetic persona responses exceeded 90 percent for most LLM model configurations. GPT-4o showed the best performance, generating synthetic responses that were strongly correlated with human responses across multiple policy domains in all three countries. While synthetic personas captured the overall direction of public opinion well, their responses were more concentrated around similar views and underrepresented the variation and more extreme opinions observed among real respondents.
These results support three recommendations for Middle East policymakers. First, they can use synthetic personas in the preliminary testing of public receptiveness to prospective policies, enabling the screening of potentially sensitive policies. Second, they can use synthetic personas to produce valuable early-stage insights even when detailed demographic data on the desired groups of people is unavailable or when resources needed to collect such data are limited; the use of synthetic personas in those cases can support faster policy cycles that are more responsive to public needs. Third, they should not rely on synthetic personas to predict the full variability or distribution of public opinion.
These results support three recommendations for Middle East policymakers. First, they can use synthetic personas in the preliminary testing of public receptiveness to prospective policies, enabling the screening of potentially sensitive policies. Second, they can use synthetic personas to produce valuable early-stage insights even when detailed demographic data on the desired groups of people is unavailable or when resources needed to collect such data are limited; the use of synthetic personas in those cases can support faster policy cycles that are more responsive to public needs. Third, they should not rely on synthetic personas to predict the full variability or distribution of public opinion.
As the Gulf becomes ever more central to the global economy and as policymaking in the region becomes increasingly complex and sensitive, policymakers should find the results of our study encouraging, for several reasons. First, the study shows that synthetic personas generated by LLMs can be a valuable tool for understanding public sentiment in the early stages of policymaking for a wide range of policies in the region. Second, it offers concrete guidance for using synthetic personas in real-world applications. Third, it demonstrates that complex prompting and rich datasets are not required to generate synthetic personas capable of simulating human policy views. Indeed, synthetic personas can be generated quickly and inexpensively using off-the-shelf LLMs. Surveys and experiments continue to be important tools for understanding how people perceive and respond to policies, but when time and funding are tight, Middle East policymakers can rely on synthetic personas.
Menu