Human-in-the-loop AI for policy communication: improving responsible gambling emails while mitigating illusion of control and emotional reactance

The digital gambling ecosystem requires combining informational accuracy with emotional management in public policy communication for Responsible Gambling (RG). After auditing the three emails that mark critical moments in a national operator's RG protocol, we compare two redesigns: a human-designed version (derived from a literature review, interviews, and focus groups) and an AI-generated version (derived from regulatory requirements). In an experiment with a mixed (between- and within-subjects) design, participants rated three responsible gambling emails, all in the same format (original, human-designed, or AI-generated), on format, content, appropriateness, and appeal (0–7), as well as emotional valence, reactance, and erroneous beliefs (winning probability, gambler's fallacy, illusion of control). Analyses treating the participant as the unit of inference (participant-level averages across the three emails within each Condition) show that both proposals (human-designed and AI-generated) outperform the original emails on perceived quality; however, strict AI–human equivalence on quality was not established under a conservative margin. Valence was comparable across proposals, while reactance was slightly higher for the AI-generated emails than for the human-designed ones. For beliefs, AI and human versions did not differ on winning probability or gambler's fallacy. By contrast, illusion of control was higher for the AI-generated proposal than for the human-designed proposal and exceeded the originals. In this dataset, the evidence is consistent with a cautious, hybrid (human-in-the-loop) approach to AI-assisted drafting in public policy communication: large language models can produce clear, non-threatening drafts, but expert editing is needed to curb illusion-of-control cues and to monitor reactance. Because AI–human equivalence was not established under the conservative margin and inferential precision is limited, any equivalence-related interpretation should be treated as provisional pending replication with larger samples and repeated-exposure designs.