Two online surveys in Japan, using the same questionnaire in 2018 and 2025, found no statistically significant difference in the item measuring intention to use relatively autonomous medical AI. Members of the public reported greater knowledge in 2025, but that did not accompany a significant rise in that intention measure. Repeated-survey study
The important qualifier is what respondents were asked to imagine: AI deciding diagnoses and treatment plans without requiring physician involvement. This was not a survey of willingness to use every form of clinician-supervised software.
The same questions did not mean the same people
The analysis compared 399 self-identified physicians and 600 members of the public in 2018 with 300 physicians and 300 members of the public in 2025. Participants came from a commercial opt-in online panel; physician status was self-reported and not independently verified.
This repeated cross-sectional design compares two samples at different times. It does not follow an individual from unfamiliarity to confidence, or show what happened to a particular person after using AI. Survey methods
The questionnaire’s intention item, usefulness judgments, self-rated knowledge and cost expectations were also separate observations. Respondents could recognise usefulness and still hesitate about handing over a clinical decision.
Among physicians, an unadjusted difference favouring AI’s potential efficiency was no longer statistically significant after age and sex adjustment. Among the public, greater self-rated knowledge coexisted with greater expectations of increased healthcare costs. The study did not establish why any individual accepted or rejected the proposed form of care.
Another Japanese study asked about use and specific applications
A separate June 2026 paper examined 200 adults surveyed in November 2025: 100 healthcare workers and 100 non-healthcare workers. It compared reported generative-AI use with acceptance of five medical applications, including assisted imaging, risk prediction and treatment recommendations. Separate use-and-acceptance study
That small convenience sample found associations between less frequent AI use and greater non-acceptance for several applications. Results varied by occupational group and application; some estimates were imprecise, with wide confidence intervals.
The finding does not overturn the repeated survey. One compares responses across years under a relatively autonomous-care definition; the other compares reported use and scenario-specific acceptance within one survey. Different questions and different populations need not produce the same pattern.
Neither design establishes that using AI causes trust. People already more accepting of AI could use it more often, while other circumstances could influence both use and acceptance. The second paper explicitly treats its results as exploratory and non-causal.
A service question needs a concrete description
For a team designing or explaining a service, these studies suggest a more precise research task than asking whether people “trust AI”. Does a clinician retain responsibility? Is the system providing information, supporting a scan interpretation or proposing treatment? What costs and data use are people being asked to accept?
Those are implications for how to frame an evaluation, not tested instructions for increasing adoption. Education alone was not experimentally evaluated as a way to change either sample’s willingness.
The repeated survey also did not establish equivalence of the questionnaire’s measurement structure across years. Its exploratory factor model had fit limitations. A non-significant change in one item should therefore not become a sweeping claim that Japanese attitudes were unchanged in every respect. Measurement and study limitations
Keep acceptance separate from competence
Our US residency-survey article concerns reported AI use and training in a different professional population. It does not measure the Japanese public’s willingness, just as willingness does not establish that a tool is accurate or its users are skilled.
The repeated Japanese study reports public programme funding; the separate study supplies another small, application-specific observation. Together they make the reader question clearer: familiarity, acceptance, actual use and effective care are distinct things to measure, with no causal shortcut between them.