The Potential of Synthetic Twin Agents for Personalized Behavioural Interventions at Scale

Personalization is increasingly emphasized in behavioural public policy, yet traditional approaches to tailoring interventions remain costly and difficult to scale. Synthetic twin agents—AI-generated counterparts of real individuals—offer a novel, scalable approach. This study evaluates whether synthetic twins generated by large language models (LLMs) can replicate human responses to personality-tailored advertisements. We recruited 373 human participants and created matched synthetic twins using four LLMs (GPT-4o, GPT-5-latest, Gemini-1.5 Flash, and Gemini-2.0 Flash). Participants evaluated advertisements for three consumer products, each tailored to Big-Five personality traits. Both human participants and their synthetic twins demonstrated personality-congruent responses for Agreeableness and Extraversion. Synthetic twins additionally showed congruence for Conscientiousness and, in some cases, Neuroticism, depending on the model and product. Trait-level comparisons showed partial alignment: Gemini-1.5 Flash most closely replicated human patterns, followed by GPT-4o and Gemini-2.0 Flash; GPT-5-latest showed the weakest alignment. Despite trait-specific divergences, aggregate correlations between human and synthetic responses were high (r = .81-.91), indicating strong consistency in relative patterns. These findings provide early evidence that prompt-based multimodal LLMs can approximate human psychological responses. Synthetic twins may serve as low-cost tools for pre-testing personalized behavioural interventions, while also raising critical ethical concerns around privacy, governance, and hyper-personalization.