'You're high risk' got more flu shots. Naming the algorithm didn't matter.

Randomised controlled trial (RCT). 90,000 people (over). three flu seasons; outcome windows 253 days (Study 1), 58 days (Study 2), 43 days (Study 3).

A trial. Chance decided who did what.

Who: Patients of Geisinger, a non-profit health system in Pennsylvania, USA, flagged by a machine-learning model as top 10% (Studies 1-2) or top 11-20% (Study 3) risk for flu and flu complications. Over 90,000 unique patients, ~130,000 observations; analysed n = 38,598, 42,481 and 39,129

• Risk note vs nothing+1.7–3.5percentage points
• Risk note vs reminder+1.1–1.4percentage points
• Saying 'algorithm'noneno clear change

More to know, from our read of the paper:

• Effects are small: +1.7 to +3.5 percentage points vs no study message, +1.1 to +1.4 vs a plain reminder; largest effect = standardised d 0.10 ('small', authors). Say 'a little', never 'a lot'
• Study 2's risk message vs plain reminder (+1.1 points) was NOT significant in the main analysis (p = .054); significant only in preregistered robustness checks. Study 3's +1.4 points vs reminder was significant (p = .016)
• All patients also got the health system's usual flu-shot marketing; 'no message' = no STUDY message
• Algorithm-vs-not comparison is a null: no significant differences in any of the 3 trials (Bayes factors favour the null, e.g. BF01 13.5 in Study 1). Say 'no clear difference', not 'people don't care about AI'. Up to ~23% may not have noticed the algorithm mention (online post-test recognition 74-82%)
• Single US health system (older, rural, less diverse); COVID-era seasons; vaccinations done elsewhere may be missed

Three randomized trials at Geisinger, a Pennsylvania health system, included over 90,000 patients that a computer model flagged as high risk for flu complications. Letters, portal messages and texts telling them they were high risk raised flu vaccination by 1.7 to 3.5 percentage points compared with no study message, and by 1.1 to 1.4 points compared with a plain flu-shot reminder. Saying a 'computer algorithm' had judged their risk made no clear difference, and adding reasons for the risk did not help either.

We read the abstract and the full text of this paper and checked every number on this page against it (17 facts, checked on 2026-10-29). The study design is named above. Cohort and other observational studies are described as linked, not proven. Found an error? Email theis@onestudyaday.com.