Public attitudes toward sharing health data for artificial intelligence: Differences by data type and sector in the Health in Central Denmark cohort

A 2024 survey of nearly 39,000 participants in Central Denmark reveals that public willingness to share health data for AI development varies significantly by data type and is substantially higher when the data is managed by public institutions rather than private ones, highlighting the critical role of institutional trust and data sensitivity in shaping acceptance.

The Gist

Imagine your health data as a giant, digital library inside your body. This library contains your medical history (text), X-rays and scans (images), recordings of your doctor's visits (audio), and your unique genetic blueprint (DNA).

Now, imagine that Artificial Intelligence (AI) is a super-smart, hungry robot that wants to read books from this library to learn how to become a better doctor. But before the robot can start reading, it needs your permission.

This study is like a massive public opinion poll asking 38,740 people in Denmark: "Would you let this robot read your library books?" But there's a twist: the researchers asked the question in two different ways to see who matters more—the who or the what.

The Two Scenarios: The "Good Neighbor" vs. The "Shopkeeper"

The researchers split the people into two groups and asked them the exact same question, but with one tiny change:

1. The "Good Neighbor" Scenario: The robot is being built by the public healthcare system (like a community library or a public hospital). Its goal is to help everyone get better.
2. The "Shopkeeper" Scenario: The robot is being built by a private company (like a tech giant or a pharmaceutical firm). Its goal is to make a profit.

The Results: What People Said

Here is how the "library owners" (the patients) reacted, using some simple analogies:

1. The "What" Matters: Some Books are More Private Than Others

People were picky about which books they were willing to lend out.

• The "Postcards" (Medical Images): People were most okay with sharing X-rays and scans. Think of these like postcards; they show a picture of a broken bone or a spot, but they don't reveal your deepest secrets. 46% said "Yes."
• The "Diary" (Medical Records): People were okay sharing text from their journals, but slightly more hesitant. 39% said "Yes."
• The "Family Heirloom" (Genetic Data): This was trickier. DNA is like a family heirloom passed down for generations; it feels very personal. 35% said "Yes."
• The "Secret Recording" (Audio): This was the hardest to share. People were very uncomfortable with the idea of a robot listening to their private conversations with doctors. Only 27% said "Yes."

2. The "Who" Matters: Trusting the Neighbor More

Regardless of what book they were sharing, people trusted the "Good Neighbor" (Public Sector) much more than the "Shopkeeper" (Private Sector).

• When the robot was working for the public, people were 12–16% more likely to say "Yes."
• When the robot was working for a private company, people were much more likely to say "No" or "I don't know."
• The Takeaway: People feel safer lending their library books to a community center than to a business that might sell them.

3. The "I Don't Know" Crowd

A huge chunk of people (about 30%) didn't know what to answer. They were like people standing at the library door, holding a book, looking confused. They weren't saying "No," but they weren't saying "Yes" either. This suggests that many people simply don't understand enough about how AI works to make a decision. They are worried about the "black box" (a mystery machine) taking their data.

Why Does This Matter?

Think of AI development as building a new, super-fast train to cure diseases. To build this train, you need tracks (data). If the public doesn't trust the train company, they won't let you lay the tracks on their land.

• The Problem: The train company (AI developers) needs data to build the train.
• The Fear: People are worried the train will be used to make money for a few rich people rather than helping the community.
• The Solution: The study suggests that if the train is built by the public (the community), with clear rules, transparent glass walls (so people can see what's happening), and strict locks on the doors (privacy), people will be much more willing to help.

The Bottom Line

If you want to build AI that helps doctors, you can't just be a "Shopkeeper" asking for data. You have to be a "Good Neighbor."

1. Be Transparent: Tell people exactly what the robot is doing with their "library books."
2. Be Public: People trust public institutions more than private ones.
3. Be Careful: Some data (like audio and DNA) is very sensitive and needs extra protection.
4. Explain It: Help people understand that the "mystery robot" is actually a tool to help them, not a spy.

In short: People are willing to share their health data for AI, but only if they trust who is holding the data and understand why they are sharing it.

Technical Summary

1. Problem Statement

The development of clinically relevant Artificial Intelligence (AI) in healthcare relies heavily on access to large, high-quality datasets. However, the deployment of AI introduces unique ethical and governance challenges beyond traditional research, including privacy concerns, algorithmic bias, "black box" opacity, and accountability issues. While public support for health data sharing is generally high in Nordic countries, there is a significant knowledge gap regarding:

• How public willingness to share data varies specifically for AI applications compared to general research.
• How attitudes differ across specific data modalities (e.g., text, images, audio, genetics).
• How the institutional context (public vs. private sector) influences trust and acceptance.

2. Methodology

Study Design & Population:

• Source: Data was drawn from the Health in Central Denmark (HICD) survey conducted in 2024.
• Sample: The study analyzed 38,740 participants (out of 136,229 invited) aged 18–92 from the Central Denmark Region.
• Weighting: To ensure representativeness of the background population (~1 million people), results were weighted using stabilized weights derived from a logistic regression model based on age, sex, education, and diabetes status.

Experimental Setup:

• Randomization: Participants were randomized into two groups to test the effect of the data recipient:
  1. Public Sector Scenario: Data shared with a public healthcare institution.
  2. Private Sector Scenario: Data shared with a private organization.
• Data Modalities: Participants were asked about their willingness to share four specific types of data for AI analysis:
  1. Electronic Health Records (text/tabular data).
  2. Medical Images (scans).
  3. Audio recordings of clinical consultations.
  4. Genetic information.
• Response Options: "Yes," "No," or "Don't know."

Statistical Analysis:

• Model: Multinomial logistic regression was used to analyze the association between the randomization group (Public vs. Private) and the response category.
• Reference Category: "Don't know" was set as the reference.
• Metrics: Odds Ratios (OR) with 95% Confidence Intervals (CI) were calculated to compare the likelihood of responding "Yes" or "No" for public vs. private sectors.

3. Key Contributions

• Scale: This is one of the largest population-based surveys on health data sharing attitudes specifically focused on AI, with nearly 40,000 respondents.
• Granularity: It is one of the first studies to explicitly differentiate between four distinct data modalities (including audio and genetics) and compare them across public and private sectors.
• Methodological Rigor: The use of randomization to isolate the variable of "data recipient" (public vs. private) allows for a causal inference regarding institutional trust, minimizing confounding factors.
• Weighted Estimates: By integrating register-based data from the entire background population, the study provides highly generalizable estimates for the region, correcting for survey non-response bias.

4. Key Results

Overall Willingness by Data Type (Public Sector Context):
Participants were most willing to share data in the following order:

1. Medical Images: 46% overall (53% in public sector).
2. Electronic Health Records (Text): 39% overall (45% in public sector).
3. Genetic Information: 35% overall (41% in public sector).
4. Audio Recordings: 27% overall (33% in public sector).

Impact of Sector (Public vs. Private):

• Public Preference: There was a consistent 12–16% higher proportion of willingness to share data with public institutions compared to private ones across all data types.
• Odds Ratios (Public vs. Private):
  • Willingness ("Yes"): OR ranged from 1.37 (Genetics) to 1.78 (EHR text). This indicates participants were significantly more likely to agree to share with public entities.
  • Refusal ("No"): OR ranged from 0.51 (EHR text) to 0.67 (Audio). This indicates significantly lower odds of refusal when the recipient is public.

Uncertainty:

• A high level of uncertainty ("Don't know") was observed across all data types (29–36%), regardless of the sector.
• Uncertainty was highest for Genetic information (36%) and Audio recordings (34%).
• Unlike some previous studies where uncertainty was concentrated in the private sector, this study found high uncertainty in both sectors, suggesting a general lack of knowledge or engagement regarding AI data use.

5. Significance and Implications

• Trust in Governance: The findings confirm that public trust in health data sharing for AI is heavily dependent on the institutional context. Public sector governance is perceived as a critical safeguard against commercial exploitation.
• Data Sensitivity Hierarchy: The public distinguishes between data types, viewing audio and genetic data as more sensitive than images or text. This suggests that tiered governance models are necessary, with stricter consent and safeguards for higher-sensitivity modalities.
• The "Black Box" Effect: The high level of uncertainty (approx. 1/3 of respondents) suggests that the complexity of AI and lack of public literacy are major barriers. The study implies that as public familiarity with AI tools (e.g., ChatGPT) increases, uncertainty may decrease, but this does not automatically translate to higher willingness without clear communication.
• Policy Recommendations:
  • Public Governance: AI development should be anchored in publicly governed infrastructures (e.g., secure hospital environments) to maintain trust.
  • Transparency: Clear communication regarding data anonymization, storage, purpose, and access is essential to reduce the "Don't know" category.
  • Regulatory Alignment: The findings support the need for frameworks like the GDPR and the European Health Data Space to provide the legal and technical safeguards (e.g., secure analysis environments) that the public demands.

Conclusion:
Public acceptance of health data sharing for AI is not monolithic; it is shaped by the perceived sensitivity of the data and the perceived motives of the data recipient. To enable responsible AI innovation, stakeholders must prioritize public benefit, ensure robust public oversight, and actively engage citizens to demystify AI technologies.