Risk nudges raised flu shots in three large US trials, AI disclosure aside

Health systems increasingly use machine-learning models to flag patients at high risk of illness, then act on that flag — often by sending a message. Two questions follow. Does telling people they are at elevated risk actually change behaviour? And does it matter whether you admit the risk estimate came from an algorithm, at a moment when public trust in medical AI is uncertain? A set of three randomised field trials published in Nature Human Behaviour on 5 October tested both, using influenza vaccination as the outcome [s1].

The trials

The studies were run in a single large US health system and preregistered, with more than 90,000 unique patients in total, and are logged on ClinicalTrials.gov under three identifiers [s1][s2]. A previously validated machine-learning algorithm identified patients at high risk for influenza and its complications [s1]. Those patients were then randomised either to receive no message or to receive one of several messages encouraging vaccination, with vaccination the primary outcome [s1].

The design let the researchers separate two things that are usually bundled together: the effect of being told you are high-risk, and the effect of how that information is framed. Some message arms disclosed that an algorithm had made the risk determination; some went further and gave personalised reasons for the algorithm's prediction, a form of what is called explainable AI; others did neither [s1].

The risk message worked, modestly

Being told you are at high risk moved behaviour. Among patients informed of their high risk, vaccination was 1.1 to 1.4 percentage points higher — a relative increase of 3.3% to 5.4% — than among patients who were simply reminded to get a vaccine [s1]. Against patients who received no message at all, the gap was larger: 1.7 to 3.5 percentage points, a relative increase of 3.3% to 14.7% [s1].

These are small absolute effects, and the article presents them as such. But nudges delivered by message cost very little per patient, so a shift of one to three percentage points across a population the size of a health system's panel translates into a meaningful number of additional vaccinations for the effort involved. The range reflects that this was three trials rather than one, run across different seasons and samples.

Disclosing the algorithm did not hurt

The more novel finding concerns framing. Vaccination was similar across message arms whether or not they disclosed that an "algorithm" was involved, and whether or not they offered the algorithm's reasons [s1]. The authors read this as evidence that, in this realistic health setting, patients were neither put off by learning an algorithm had flagged them nor won over by it — they were, as the paper puts it, neither averse to nor appreciative of the algorithm's use or its explainability [s1].

That matters because laboratory studies of attitudes toward medical AI have been mixed, with some suggesting people recoil from algorithmic involvement in their care. A field trial at this scale, measuring actual vaccinations rather than stated attitudes, is a stronger test of whether that aversion shows up when something is at stake. Here it did not.

What the study does not show

Several limits bound the result. The trials were conducted in one US health system, and the patients were those its algorithm identified as high-risk, so the effects are not necessarily transferable to general populations or to other health systems' models [s1]. The outcome was influenza vaccination specifically; it does not follow that risk messaging or AI disclosure would behave the same way for a more consequential or more contested intervention. And a null result on disclosure — patients not reacting to the word "algorithm" — is a finding about this context, not a guarantee that transparency is always neutral.

There is also a commercial thread to note: the health system paid an outside firm to generate the risk scores and the reasons used in the studies, a relationship the authors disclose [s1]. The work was partially funded by the US National Institute on Aging [s1].

Why it is useful anyway

The practical value is in the pairing of the two findings. A health system now has field evidence that a risk-based nudge produces a small but real increase in flu vaccination, and that it can be honest about using an algorithm to target that nudge without paying a penalty in uptake. For organisations weighing whether to hide or disclose the machinery behind personalised health messages, the second point removes one argument for opacity.

This article describes research findings and is not medical advice.

Sources

• [s1] Algorithm-assisted personalized risk communication to encourage flu vaccination in the USA: three randomized field trials. Nature Human Behaviour, 5 October 2026. https://doi.org/10.1038/s41562-026-02603-4
• [s2] Encouraging Flu Vaccination Among High-Risk Patients Identified by ML (NCT04323137). ClinicalTrials.gov, first posted 26 March 2020. https://clinicaltrials.gov/study/NCT04323137