📊 statistics

Public acceptance of artificial intelligence in health care: a cumulative hierarchy in a 16-item instrument, German–Italian measurement equivalence, and a five-item short form

This study demonstrates that public acceptance of artificial intelligence in healthcare follows a cumulative hierarchy ordered by clinical consequence, remains invariant across German- and Italian-speaking populations, and can be effectively captured by a five-item short form despite the lack of a uniquely determined item set.

Original authors: Dietmar Ausserhofer, Christian J. Wiedermann, Verena Barbieri, Nicoletta Colletti, Stefano Lombardo, Timon Gärtner, Giuliano Piccoliori, Doris Hager von Strobele-Prainsack, Wolfgang Wiedermann

Published 2026-09-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dietmar Ausserhofer, Christian J. Wiedermann, Verena Barbieri, Nicoletta Colletti, Stefano Lombardo, Timon Gärtner, Giuliano Piccoliori, Doris Hager von Strobele-Prainsack, Wolfgang Wiedermann

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine standing in a hospital waiting room, watching a screen display a doctor's diagnosis generated by a computer program. For some, this feels like a helpful assistant; for others, it feels like a dangerous intrusion. As artificial intelligence moves from research labs into real hospitals, doctors, and pharmacies, a crucial question arises: do the people these systems are meant to serve actually accept them? The answer is rarely a simple "yes" or "no" for every single task. Instead, public opinion often forms a pattern. People might be comfortable with a machine scheduling an appointment but deeply uneasy about a machine deciding on a life-or-death treatment. Understanding this pattern is vital because it tells us not just what people think, but how their trust is built or broken across different types of medical work.

A team of researchers in South Tyrol, a bilingual region in northern Italy where German and Italian speakers live side by side, set out to map this landscape of trust. They wanted to know if public acceptance of artificial intelligence in healthcare follows a predictable order, if that order holds true for both language groups despite cultural differences, and if they could create a shorter, simpler way to measure it. They surveyed 901 adults from the general population, asking them to rate sixteen different ways artificial intelligence could be used in medicine. These uses ranged from administrative tasks, like organizing patient records and sending medication reminders, to high-stakes clinical decisions, like supporting a doctor during a medical emergency or helping to diagnose a complex illness.

The researchers found that the public's acceptance does indeed follow a clear, cumulative ladder. At the bottom of the ladder, where acceptance is highest, are the administrative tasks. Nearly eighty percent of the people surveyed said they were comfortable with AI sending automated reminders to take medicine. A little lower, but still widely accepted, were tasks like scheduling appointments and managing electronic health records. As the tasks became more clinical and involved more direct patient care, acceptance began to drop. People were less comfortable with AI helping to analyze clinical data or support a diagnosis. At the very top of the ladder, where acceptance is lowest, are the time-critical, high-pressure decisions. Only about a third of respondents said they would accept AI support during a medical emergency or triage, and even fewer accepted it for making urgent decisions. This hierarchy suggests that the public does not view artificial intelligence as a single block of technology; rather, they judge it based on how much clinical consequence the specific task carries. The more a task feels like a direct medical decision, the more hesitant people are to let a machine handle it.

Crucially, this ladder of acceptance was identical for both the German-speaking and Italian-speaking populations in the study. Even though the survey questions were translated and adapted to fit the cultural nuances of each language group, the order in which people ranked the tasks remained exactly the same. This finding is significant because it proves that the way people evaluate these technologies is not just a matter of language or local culture, but a shared human response to the nature of the medical task itself. It means that health systems serving multiple languages can use the same tool to measure trust without needing to treat the groups as fundamentally different.

The researchers also discovered that they could shrink the survey from sixteen questions down to just five without losing the ability to see this pattern. By carefully selecting five specific questions that covered the full range from easy administrative tasks to difficult emergency decisions, they created a short form that was just as accurate as the long one. This shorter version is particularly useful for busy surveys where time is limited, allowing researchers to quickly gauge public sentiment without overwhelming respondents. However, they noted that while the two most extreme questions—one about medication reminders and one about urgent decisions—were essential anchors, the three questions in the middle could vary slightly depending on the specific context, as long as the overall range of difficulty was preserved.

The study also looked deeper into the types of people holding these views. While most people's opinions followed the general ladder, about thirty percent of the respondents formed a distinct group. These individuals were comfortable with AI handling administrative work but drew a hard, categorical line against letting it touch any clinical decisions. For them, the difference between a machine organizing a file and a machine helping a doctor was not a matter of degree, but of kind. They saw the two realms as separate, accepting one while rejecting the other entirely. This suggests that for a significant portion of the population, the boundary between administrative support and clinical care is a bright line that cannot be crossed by technology.

Finally, the researchers examined how age and digital experience influenced these views. In a simple count of the data, older adults appeared more likely to be found at both extremes: they were overrepresented among those who rejected all AI and among those who accepted it broadly. However, when the researchers adjusted for how much people actually used technology in their daily lives, this pattern changed. The apparent polarization among older adults was largely explained by their varying levels of digital experience. Those who used technology frequently were less likely to reject AI, regardless of their age. This indicates that familiarity with technology, rather than age itself, is a stronger driver of acceptance.

The study concludes that public trust in healthcare artificial intelligence is not a chaotic mix of random opinions, but a structured hierarchy based on the clinical weight of the task. This structure is robust, holding true across different languages and cultures, and can be measured effectively with a very short set of questions. While the findings offer a clear map of where the public stands today, the researchers caution that this map applies to the general population. Whether doctors and nurses, who work with these systems every day, follow the same ladder of trust remains a question for future study. For now, the data provides a solid foundation for understanding how to introduce artificial intelligence into healthcare in a way that respects the public's boundaries.

Problem and Motivation
Public surveys regarding artificial intelligence (AI) in health care typically report acceptance rates for specific applications in isolation. This descriptive approach fails to determine whether acceptance forms a systematic, ordered cumulative structure, whether such a structure holds across different language groups within a single health system, or whether the construct can be captured efficiently by a short form. Furthermore, existing instruments often lack validation in multilingual contexts where translations are culturally adapted rather than literal, raising questions about measurement equivalence. This study addresses these gaps by analyzing the internal structure of a 16-item instrument measuring public acceptance of AI in health care, administered in the bilingual (German- and Italian-speaking) population of South Tyrol, Italy.

Methodology
The study utilized data from a probabilistic population panel in South Tyrol surveyed in February–March 2026 (N = 901). Respondents rated the acceptability of 16 distinct AI applications ranging from administrative functions to time-critical clinical decisions. Uncertainty responses were scored as non-acceptance.

The analysis employed nonparametric item response theory, specifically Mokken scale analysis, to assess:

Scalability and Dimensionality: Using Loevinger coefficients ( $H$ , $H_i$ , $H_{ij}$ ) to determine if items form a unidimensional scale.
Monotonicity and Invariant Item Ordering (IIO): Checking if the probability of a positive response increases monotonically with the latent trait and if the ordering of item difficulties is consistent across respondents ( $H_T$ coefficient).
Local Independence: Assessed via conditional association and Yen's $Q_3$ residual correlations.
Measurement Equivalence: Differential Item Functioning (DIF) was evaluated using logistic regression on the sum score, comparing German and Italian versions. Effect sizes (increment in McFadden pseudo- $R^2$ ) were prioritized over significance testing, with a relevance threshold of 0.035.
Short Form Derivation: An exhaustive search of all 5- and 6-item subsets was conducted. Selection criteria included scalability, person ordering ( $H_T$ ), local independence, correlation with the full 16-item score, and coverage of the difficulty range. The selection rule was cross-validated over 200 random split halves (400 selection-evaluation pairs).
Latent Class Analysis (LCA): Models with 2–6 classes were fitted to explore whether respondents cluster into distinct acceptance types. Associations with covariates were estimated using a three-step procedure with BCH correction and sandwich standard errors, supplemented by nonparametric bootstrapping to account for measurement model uncertainty.

Key Results

Cumulative Hierarchy: The 16 applications formed a single moderately strong scale ( $H = 0.491$ ). The hierarchy is ordered by clinical consequence: administrative applications (e.g., medication reminders, appointment scheduling) are most accepted, while time-critical clinical decisions (e.g., emergency triage, urgent decisions) are least accepted.
Invariant Item Ordering: While the scale supports group comparisons, invariant item ordering across individual respondents was not supported for the full 16-item instrument ( $H_T = 0.267$ ). This indicates that while the scale orders groups reliably, it cannot precisely locate individuals based on the item hierarchy alone.
Measurement Equivalence: The ordering of item difficulties was nearly identical across language groups (Spearman $\rho \ge 0.962$ ). No item exhibited relevant DIF between the German and Italian versions (largest pseudo- $R^2$ increment = 0.0070), even for items where the Italian version involved culturally adapted shifts in scope or omitted qualifiers.
Short Form: A five-item short form was derived, achieving $H = 0.597$ and significantly better person ordering ( $H_T = 0.507$ , 97.7th percentile) compared to the full instrument. It correlated 0.903 with the full score. However, the specific composition of the short form was not uniquely determined; the selection rule yielded 34 different forms across cross-validation splits. The adopted form includes "automated medication reminders" and "support in urgent decisions" as anchors, with intermediate items selected based on content coverage rather than unique psychometric superiority.
Acceptance Typology: Latent class analysis identified four classes. A distinct "administratively oriented" class (29.7% of respondents) was found, which accepts administrative applications at high levels but rejects clinically consequential applications at levels similar to the "rejecting" class. This suggests a categorical boundary between administrative and clinical AI for a significant minority, rather than a purely gradual difference in intensity.
Covariates: Language group and area of residence were not associated with class membership. Age showed a complex relationship: while older respondents were descriptively over-represented at the extremes of the typology, this association reversed after adjusting for digital experience (AI use and technology affinity), suggesting the descriptive pattern was a compositional effect.

Significance and Claims
The paper claims that public acceptance of AI in health care is not a collection of unrelated judgments but follows a cumulative hierarchy ordered by the clinical consequence of the application. This hierarchy is invariant across German- and Italian-speaking populations, supporting the use of culturally adapted parallel versions of the instrument without stratified scoring.

The study demonstrates that a five-item short form can reproduce this hierarchy with improved person ordering, making it suitable for settings where the full instrument is impractical. However, the authors modestly note that the specific items in the short form are not uniquely determined by the data; the anchors are fixed, but intermediate items are interchangeable based on content needs.

The findings also challenge the assumption that acceptance is purely a matter of degree; for approximately 30% of the population, the distinction between administrative and clinical AI is categorical. Finally, the study highlights that descriptive age differences in acceptance may be confounded by digital experience, reversing direction when adjusted.

The authors explicitly state limitations: the hierarchy's generalizability to health professionals or other health systems remains untested; the short form composition requires external validation; and the latent class solution is exploratory, as formal model selection criteria did not converge on a specific number of classes.

More like this