Conversational AI for Adult Flu Vaccination Decisions

In a prior randomized trial by the same investigator, a similar intervention (evidence cards followed by an AI-guided motivational-interviewing conversation) raised MMR vaccination intention among hesitant parents by about one point on a 1-7 scale relative to an active control (Forman 2025, PsyArXiv, doi:10.31234/osf.io/5kteh_v1). This study tests whether the effect transfers to a different decision and population: adult seasonal flu vaccination.

Design: two-arm, parallel, 1:1 randomized trial conducted entirely online. Participants are US adults on Prolific (approval rate ≥ 99%, at least 25 previous submissions), desktop or mobile. A five-item health-plans screener masks the eligibility question; eligible participants answer "Probably", "Unsure", or "Probably not" to getting a flu vaccine in the next few months. Participants then answer the baseline intention item; those answering 7 are screened out before randomization. Randomization is by minimization within screener stratum (probably / unsure / probably not), with random tie-breaking. Up to 220 are randomized; dropouts are not replaced.

Intervention: a study version of Calypta Health's public flu vaccination guide: an intro card and 12 evidence cards (including an interactive simulation), followed by a conversation with an AI model (a Claude model accessed through Amazon Bedrock; the model and its instructions are fixed for the study's duration). Control: a study-only guide on everyday accident avoidance (mostly driving, plus falls and smoke and carbon-monoxide alarms) with the same structure, card count, minimum reading time, and conversation rules. In both arms, per-card minimum times (10 seconds for the intro card, 20 seconds for each other card) and a minimum conversation length (at least 3 rounds and 100 engagement points) apply. After the post-session question, participants in either arm who answer 6 or 7 are offered the same optional flu-vaccination plan card (where and which day). Because the card is offered only after a 6 or 7, more flu-arm participants may see it, so the vaccination comparison includes any effect of the card.

Follow-up: check-ins at about 2 and 8 weeks, each a separate Prolific study open to everyone randomized, ask about flu vaccination since the session, intention, and (as a negative-control outcome) dental cleanings or checkups.

Analysis: the primary outcome is analyzed by ANCOVA adjusting for baseline intention and screener stratum, with HC3 standard errors, two-sided α = .05. The analysis set is randomized participants who completed the session and met the engagement rule; the primary population excludes sessions that an AI reviewer links as one person using two accounts, sensitivity analyses include everyone or remove sessions flagged as low quality, and a further analysis is limited to participants with a baseline intention of 5 or below. Analysis code for the primary outcome is written and frozen before arm labels are unblinded; check-in outcomes are analyzed as the preregistered plan specifies once each wave closes, with any deviation reported. A full analysis plan is preregistered on OSF and linked from this record.

A feasibility pilot (20 randomized, 2026-10-15), under the same IRB approval, preceded main enrollment to test study mechanics and a next-day follow-up; pilot data are not presented as results of this study or combined with main-study data, and pilot participants are excluded from the main study.

Conflict of interest: the investigator founded Calypta Systems, Inc. and is its CEO, sole director, and majority shareholder. Calypta makes the flu guide and funds the study. Results will be made public regardless of outcome.