Paper: Global analysis of country-level factors associated with chatbot usage for health
Authors: Philipp Schoenegger, Beatriz Costa-Gomes, Pavel Tolmachev, and colleagues
Published in: Nature Health, July 2026
Estimated reading time: 5–7 minutes
Source: Read the original paper in Nature Health (https://www.nature.com/articles/s44360-026-00174-2)
Overview
People increasingly ask general-purpose AI chatbots about symptoms, treatments, mental wellbeing, medical forms, and how to access healthcare. However, chatbot use is unlikely to be the same everywhere. A person’s local healthcare system, access to technology, economic conditions, and trust in medical institutions may all shape how they use AI.
This paper investigates those differences on a global scale. The researchers analysed 1.7 million health-related Microsoft Copilot conversations from 109 countries and regions, collected between January and March 2026. Their central finding is that two separate factors appear to be at work:
- Trust in hospitals is associated with how often people use Copilot for health.
- A country’s development level and healthcare structure are associated with what people ask the chatbot.
This distinction—between the amount of health-related use and the type of use—is the paper’s most important contribution.
What Did the Researchers Want to Know?
The study addresses two main questions.
First, why do health questions represent a larger share of chatbot activity in some countries than in others? The paper calls this health conversation intensity.
Second, how does the subject of those conversations vary between countries? The authors call this intent composition. For example, users might seek general health education, ask about symptoms, request emotional support, or need help with medical paperwork.
This separation matters. A country could have a relatively high proportion of health-related conversations but still have a very different mix of questions from another country with a similar level of use.
Data and Method
The dataset contained 1.7 million health-related conversations from consumer Copilot accounts. Enterprise, educational, and commercial accounts were excluded. The researchers included only countries or regions with at least 1,000 health conversations during the three-month study period.
To protect privacy, the original conversations were stripped of personally identifiable information. A language model then converted each conversation into a short English description of its topic and purpose. Researchers did not directly read the original messages.
Each conversation was assigned to one of eight major intent categories. These included areas such as:
- General health information and education
- Research or academic support
- Symptom-related questions
- Fitness and lifestyle
- Emotional wellbeing
- Medical paperwork
- Healthcare navigation
The classifier used for this task had previously achieved 84% exact-match accuracy when compared with classifications made by clinicians. The eight selected categories covered more than 95% of the health conversations.
The researchers measured chatbot use in two ways:
- Intensity: the percentage of all Copilot conversations in a country that were health-related.
- Composition: the percentage of health conversations belonging to each intent category.
They then compared these measurements with country-level indicators such as GDP per person, internet access, population age, physician availability, health spending, universal health coverage, government AI readiness, social protection, and public trust in hospitals and medical professionals. Because complete information was not available for every country, the main regression analysis used 93 of the 109 countries.
Importantly, this was an observational, country-level study. The statistical analysis identifies associations, not cause-and-effect relationships.
Main Findings
1. Health questions form a meaningful part of chatbot use
Across the countries studied, health conversations represented an average of 8.7% of all Copilot conversations. The figure ranged from approximately 4.4% to 15.35%.
In practical terms, close to one in every eleven conversations involved health. This supports the view that health is already an important use case for general-purpose conversational AI, even though Copilot is not designed only as a medical system.
2. Lower confidence in hospitals was associated with greater health-related use
The strongest individual predictor of health conversation intensity was public confidence in hospitals. Countries where fewer people reported confidence in hospitals tended to have a larger proportion of Copilot conversations about health.
The correlation was moderately negative at r = −0.41. After accounting for economic, demographic, healthcare, and institutional factors, hospital confidence remained the strongest predictor in the main model.
The authors estimate that a decrease of roughly 13 percentage points in hospital confidence was associated with about a one-percentage-point increase in health conversation intensity. Relative to the average intensity of 8.7%, that is an increase of approximately 11%.
This finding may suggest that people turn to AI when they are less confident in formal healthcare institutions. However, the evidence for this relationship is described as moderate, because the result was more sensitive to changes in the statistical specification than the study’s findings about conversation topics.
It would also be incorrect to conclude that individuals who distrust hospitals are necessarily the same individuals who use chatbots for health. The study only compares national averages.
3. Economic development affected what people asked
The composition of health conversations followed a noticeable development pattern.
In lower-income countries with younger populations, conversations were more likely to involve broad health information, education, and academic research. These categories accounted for more than half of the health-related questions in such settings.
In wealthier countries with older populations, users were more likely to ask specific and personally actionable questions. These included understanding symptoms and navigating healthcare services, such as finding the appropriate service or making sense of a referral process.
Healthcare navigation was the most predictable category in the analysis. Its relationship with GDP, the proportion of older people, and national AI readiness remained statistically robust after the authors corrected for multiple comparisons.
A reasonable interpretation is that chatbot use reflects a user’s surrounding environment. Where access to formal healthcare is limited, AI may serve as a general source of health knowledge. Where healthcare systems are more developed, users may employ AI to manage specific interactions with those systems.
4. Structured healthcare systems produced more paperwork questions
Medical paperwork did not follow the same economic pattern as most other categories. Instead, it was strongly associated with the Universal Health Coverage index.
A one-standard-deviation increase in universal health coverage was associated with a four-percentage-point increase in the share of conversations about medical paperwork. Since paperwork represented about 7% of health conversations on average, this is a substantial difference.
One possible explanation is that better-organized healthcare systems create more administrative tasks, such as completing forms, interpreting documents, handling referrals, or understanding insurance processes. AI may therefore help people manage the bureaucracy created by formal healthcare.
The authors are careful about this interpretation. More paperwork questions could also indicate greater digital integration or stronger patient engagement, rather than simply an excessive administrative burden.
Why This Paper Matters for Computer Science
From a computer science perspective, the study shows that chatbot behaviour cannot be understood by looking only at aggregate usage statistics. Counting health conversations tells us how much a system is being used, but intent classification reveals how it is being used and what needs it may be serving.
The findings also have direct implications for system design. A health chatbot used in a low-trust environment may need strong safeguards against self-diagnosis, clearer explanations of uncertainty, and reliable handover to human professionals. Systems used for healthcare navigation may need accurate, locally relevant information about services and referral pathways.
The study also illustrates a privacy-conscious analysis pipeline: personal information was removed, conversations were summarized automatically, and analysis was conducted using country-level aggregates. At the same time, the 84% classifier accuracy leaves room for error. Performance could vary across languages or countries, and the researchers could not test this using the original multilingual conversations.
Important Limitations
Several limitations should shape how the results are interpreted:
- The study is cross-sectional and cannot establish causality.
- Copilot users are not representative of entire national populations.
- Intensity is measured as a share of Copilot activity, not as chatbot use per person.
- Only one AI platform was studied.
- Some trust measurements came from 2018, while the conversations were collected in 2026.
- Country-level patterns cannot be converted into claims about individual behaviour.
- The underlying conversation data and analysis code are not publicly available, limiting independent reproduction.
- All authors were Microsoft employees, and Microsoft operates the platform being studied.
These limitations do not invalidate the findings, but they mean the paper should be seen as an early global picture rather than a final explanation of why people use AI for healthcare.
Conclusion
The paper’s clearest message is that health-related chatbot use has two different dimensions. Institutional trust is associated with how much AI is used for health, while economic development and healthcare structure are associated with what users ask.
Conversational AI may help people understand health information, overcome access barriers, and navigate complicated healthcare systems. Yet it could also become a substitute for professional care in places where trust is already weak. Future studies will need data from multiple platforms, observations over longer periods, individual-level evidence, and connections to real health outcomes.
For researchers and developers, the practical lesson is simple: health chatbots should not be treated as culturally neutral systems. Their role changes across countries, and safe design must account for local institutions, languages, levels of trust, and healthcare needs.
