🤖 AI

LLMs for health: Perceived benefits, risks, intention to use AI chatbots, and willingness to self-disclose across sensitive health topics

Based on a study of 1,388 Dutch participants, this paper finds that the intention to use AI chatbots for health inquiries and the willingness to self-disclose sensitive information are primarily driven by perceived benefits, perceived risks, and individual characteristics rather than by the specific type or sensitivity of the health topic discussed.

Original authors: Gwenn Beets, Anniek Jansen, Saar Hommes, Ruben D. Vromans, Leonie Westerbeek, Supraja Sankaran, Julia C. M. van Weert, Emiel J. Krahmer, Nadine Bol

Published 2026-07-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gwenn Beets, Anniek Jansen, Saar Hommes, Ruben D. Vromans, Leonie Westerbeek, Supraja Sankaran, Julia C. M. van Weert, Emiel J. Krahmer, Nadine Bol

Original paper licensed under CC BY 4.0 (http://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 your health questions are like a giant, messy backpack full of different items. Some are light and easy to talk about, like a stubbed toe or a common cold (physical, low-sensitivity). Others are heavy, awkward, and maybe a bit embarrassing to pull out in public, like anxiety or a weird rash (psychological, high-sensitivity).

For a long time, people thought that if you asked an AI chatbot about those heavy, awkward items, you'd be more likely to spill your secrets because the robot doesn't judge. But this new study, which asked 1,388 real people in the Netherlands to imagine using an AI for health, suggests that's not quite how the backpack works.

Here is what the researchers found, using the tools of a giant digital experiment.

The Big Trade-Off: The "Good Stuff" vs. The "Scary Stuff"

The main discovery is that when people decide whether to talk to an AI about their health, they are playing a mental game of tug-of-war. On one side of the rope, they have the Perceived Benefits (the "Good Stuff"). On the other side, they have the Perceived Risks (the "Scary Stuff").

The study measured this with a 1-to-5 scale. On average, people saw the "Good Stuff" as a 3.65 and the "Scary Stuff" as a 3.54. It was a pretty even match!

But here is the kicker: The "Good Stuff" pulled the rope much harder.

If a person thought the AI would be helpful (high benefits), they were much more likely to say, "Yes, I'll use it!" and "Yes, I'll tell it my secrets!"
If a person thought the AI might leak their data or give bad advice (high risks), they were less likely to use it.

The paper suggests that the "Good Stuff" is the heavyweight champion here. Even though people were worried about risks, the promise of getting quick answers or unlimited questions was the main driver for wanting to use the chatbot.

The "Topic" Myth: It Doesn't Matter What You Ask

You might think, "Well, I'd definitely tell a robot about my headache, but I'd never tell it about my depression."

The researchers tested this by giving people two types of scenarios:

Physical vs. Psychological: Is the problem a broken bone or a bad mood?
Low vs. High Sensitivity: Is the problem a minor itch or a deep, stigmatized fear?

The surprise: The type of problem barely changed anything.
Whether the topic was physical or psychological, or low-sensitivity or high-sensitivity, it did not significantly change how much people thought the AI was risky or beneficial. The only tiny difference the study found was that people were slightly more willing to use the AI for low-sensitivity topics (like a mild headache) than high-sensitivity ones (like a serious mental health crisis), but the difference was so small it was almost invisible.

So, the idea that "AI is great for secrets because it's non-judgmental" wasn't the main factor here. The topic itself wasn't the boss; the person's own feelings about benefits and risks were.

Who is the Robot Friend? (It's About You, Not the Topic)

If the topic doesn't matter, what does? The study found that who you are matters a lot more than what you ask.

Think of the AI chatbot as a new tool in a toolbox. Some people grab it immediately; others stare at it suspiciously. Here is who grabs it and who puts it back:

The "I've Used It Before" Crew: People who had already used an AI chatbot were much more likely to use it again and share their health data. If you've played with the tool before, you trust it more.
The "I Don't Know What It Is" Crew: People who had never heard of AI chatbots, or heard of them but never used them, were much less likely to use them. They were like people staring at a new smartphone, afraid to touch the screen.
The "Super-Researchers": People who liked to hunt for information themselves (a style called "monitoring") were more likely to use the AI. But if you preferred to just ask your doctor everything and didn't like searching on your own, you were less likely to use the bot.
The "High-Education" Group: Interestingly, people with higher education levels actually saw more risks. They were the ones thinking, "Wait, could this leak my data?" or "Is this advice accurate?" They were the cautious skeptics.
The "Trust" Factor: People who generally trusted institutions (like the government or healthcare systems) saw more benefits in the AI. If you trust the system, you trust the robot inside it.

The Numbers Behind the Curtain

The study didn't just guess; they measured it.

1,388 people participated, which is a huge group representing the Dutch population.
They looked at 16 different health scenarios (ranging from mild to severe, physical to psychological).
The link between "thinking it's useful" and "wanting to use it" was strong (a score of 0.51 on a statistical scale).
The link between "thinking it's risky" and "not wanting to use it" was there too, but weaker (a score of -0.14).

What the Study Doesn't Say

It's important to know what this study didn't prove.

It did not prove that AI chatbots are actually safe or dangerous in real life. It only measured what people thought about them.
It did not find that age or gender made a big difference. Unlike some other studies, this one found that being young or old, or being a man or a woman, didn't change how people felt about the AI.
It did not say that the AI is a "solved" problem. The authors suggest that because the study was based on imagination (people imagining using the bot), real-life behavior might be different.

The Bottom Line

The study suggests that if you want people to use AI for their health, you can't just say, "It's great for your secrets!" The topic (whether it's a headache or a heartache) isn't the magic key.

Instead, the key is trust and experience. If people have used the tool before, if they feel it's useful, and if they don't feel like it's a ticking time bomb of risks, they will use it. But if they've never touched it, or if they are the type of person who worries about privacy and likes to double-check everything with a human doctor, they will likely keep the AI on the shelf.

The researchers suggest that as AI becomes more common, these feelings might change, but for now, the "Good Stuff" vs. "Scary Stuff" tug-of-war is the real story.

Technical Summary: LLMs for Health – Perceived Benefits, Risks, Intention to Use, and Willingness to Self-Disclose

Problem Statement
The rapid adoption of Large Language Model (LLM) chatbots for health-related inquiries has shifted user behavior from static web searches to interactive dialogues. While search engines remain primary sources for health information, a significant portion of the population (37% in the Netherlands) now utilizes AI chatbots. However, fundamental questions remain regarding public awareness of the trade-offs involved: specifically, how perceived benefits and risks influence the intention to use these tools and the willingness to self-disclose sensitive health information. Furthermore, it is unclear whether the nature of the health topic (physical vs. psychological) and its sensitivity level (low vs. high) significantly alter these perceptions and behaviors, or if individual characteristics (e.g., AI literacy, demographics, coping styles) play a more dominant role. Existing research often focuses on single healthcare contexts or stigmatized topics in isolation, lacking a systematic comparison across diverse health conditions and sensitivity levels.

Methodology
The study employed a large-scale online experiment with a representative Dutch sample ( $N = 1,388$ ) recruited via the LISS panel. The design was a mixed factorial $2 \times 2$ structure:

Between-subjects: Topic type (Physical vs. Psychological conditions).
Within-subjects: Topic sensitivity (Low vs. High).

Participants were presented with sixteen scenarios derived from a pretest of 24 health conditions, ensuring conditions were comparable in severity but distinct in sensitivity. Each participant viewed two scenarios (one low-sensitivity, one high-sensitivity) within their assigned topic type (physical or psychological).

Measures and Analysis

Dependent Variables: Perceived benefits, perceived risks, intention to use, and willingness to self-disclose. These were measured using Likert scales adapted from established taxonomies (e.g., Antes et al., 2021; Weidinger et al., 2022).
Independent Variables:
- Experimental: Topic type and sensitivity.
- Individual Characteristics: Age, gender, education, AI literacy, eHealth literacy, political orientation, trust in institutions, coping styles (monitoring), and prior experience with AI chatbots and health services.
Statistical Approach: The authors utilized Linear Mixed Effects Models (LMM) and MANOVA to test hypotheses and research questions. Models accounted for random intercepts for participants and experimental conditions. Standardized estimates were reported, and alpha corrections were applied to account for multiple testing ( $p < .025$ for hypotheses, $p < .0125$ for research questions).

Key Results

Benefits and Risks as Primary Drivers: Consistent with the Privacy Calculus and Health Belief Model, perceived benefits were positively associated with both the intention to use AI chatbots and the willingness to self-disclose. Conversely, perceived risks were negatively associated with these outcomes. Notably, the effect sizes for benefits were larger than those for risks, suggesting immediate perceived utility may outweigh potential long-term risks in decision-making.
Limited Impact of Topic Type and Sensitivity: While a multivariate effect of topic type was observed, univariate analyses revealed that topic type and sensitivity had negligible effects on perceived benefits, risks, and willingness to self-disclose. The only significant finding was a small difference in intention to use: participants reported slightly higher intentions for low-sensitive topics compared to high-sensitive topics.
Significance of Individual Characteristics: Individual factors were strongly associated with outcomes:
- AI Experience: Participants with prior experience using AI chatbots reported higher intentions to use and greater willingness to self-disclose compared to those with no experience.
- Literacy and Coping: Higher AI literacy correlated with greater willingness to self-disclose. Conversely, higher eHealth literacy and a coping style focused on asking specialists many questions were associated with lower willingness to self-disclose and lower intention to use chatbots.
- Demographics and Trust: Higher education levels were associated with perceiving greater risks. Trust in institutions positively correlated with perceived benefits. Political orientation (left vs. right) also influenced risk perception.

Key Contributions

Systematic Comparison: The study provides a rare systematic comparison of physical versus psychological health topics and varying levels of sensitivity within a single experimental framework, moving beyond isolated case studies.
Nuanced View of Individual Differences: By integrating a wide array of individual characteristics (from political orientation to specific coping styles), the paper offers a granular understanding of who is likely to adopt health-related AI and under what conditions.
Empirical Validation of Trade-off Models: The findings empirically support the application of the Privacy Calculus and Health Belief Model to the context of LLMs in healthcare, confirming that the benefit-risk trade-off is the central mechanism driving user behavior.

Significance and Claims
The paper claims that intentions to use AI chatbots for health and the willingness to self-disclose sensitive information are primarily driven by perceived benefits and risks and individual characteristics, rather than the specific type or sensitivity of the health topic discussed.

The authors position this work as a foundational step for researchers, policymakers, and developers. They suggest that because topic sensitivity has a minimal direct impact on usage intentions, efforts to improve AI adoption in health should focus on:

Enhancing the perceived benefits of AI tools.
Addressing specific risk perceptions, particularly among high-educated groups and those with high eHealth literacy.
Tailoring interventions based on individual user profiles (e.g., AI literacy levels and coping styles) rather than assuming a uniform response to different health topics.

The study concludes with a modest note that while the findings are robust within the experimental context, real-world trade-offs may differ due to contextual factors not captured in a survey setting. The authors emphasize the need for future research to examine actual conversations and real-life decision-making processes.