The Effect of AI-based Linguistically Tailored Health Messages on Plant-Based Lunch Choices of Adolescents: A Randomised Controlled Trial

Background Adolescents are a notoriously hard-to-engage audience in digital health communication. Many strategies have aimed to increase communication effectiveness by personalising what is said, but have largely neglected how it is said. Addressing this gap, we introduce an innovative approach to personalised health communication: automated AI-based linguistic tailoring. Using large language model style transfer, we mimic adolescents’ individual communication style to test if linguistically tailored health messages promote attitude change and healthier dietary behaviour. Methods We conducted a preregistered 2-week randomised controlled trial with 587 adolescents (aged 13–16) in the UK and Ireland. Participants were allocated to one of four arms: no health messages (passive control), non-tailored health messages (active control), tailored health messages, or both message types presented in counterbalanced order in weeks 1 and 2. This enabled between-subject and within-subject comparisons. The primary outcome was plant-based lunch choice (yes/no); secondary outcomes assessed evaluations of the messages and their source. Findings Adolescents receiving health messages (tailored or not) increasingly chose plant-based lunches more often over time than those receiving no messages (OR = 1.06, 95% CrI [1.02, 1.09]; pd = 1.000; BF₁₀ = 35.39). This effect was similar for tailored and non-tailored conditions (OR = 1.01, 95% CrI [0.98, 1.05]; pd = .731; BF₁₀ = 0.12). However, exploratory analyses indicated higher ratings of relevance (BF₁₀ = 3.50), personalisation (BF₁₀ = 4.93), and source liking (BF₁₀ = 3.38) for linguistic tailoring in the second week of the intervention. Interpretation Fact-based dietary health messages effectively promoted plant-based lunch choices, although this behavioural effect seemed driven more by content than by style. However, with repeated exposure, linguistic tailoring enhanced several message and source evaluations, indicating promise for future AI-driven personalised health campaigns and youth communication.