Trust in Generative AI for Health Information Consumption and the Effect of Learned Dependency: An Experimental Investigation

This new research from arXiv tackles a critical, often-overlooked psychological side effect of using generative AI for health information: learned dependency. The study moves beyond simple accuracy metrics to ask a more profound question: as users rely on GenAI for health answers, does their ability to critically evaluate that information degrade?

What Happened

The experiment, detailed in the paper (arXiv:2606.20605v2), investigates how repeated reliance on GenAI for health queries alters a user’s trust calibration—the ability to match one’s trust to the actual reliability of the source. The core hypothesis is that as users become dependent on the AI’s convenient, conversational answers, they stop engaging in the cognitive effort required to verify or question those outputs. This creates a feedback loop: the more you use it, the less you question it, and the more vulnerable you become to errors or hallucinations. The study’s design likely compared a control group (using traditional search or static information) against a group that built a habit of consulting GenAI, measuring shifts in both trust levels and critical evaluation skills over time.

Why It Matters

This is not just an academic curiosity. Health information is high-stakes. A hallucinated drug interaction or a misinterpreted symptom summary can have real-world consequences. The research highlights a dangerous paradox: GenAI is designed to be helpful and authoritative in tone, which can mask its unreliability. The concept of learned dependency suggests that even if a model’s accuracy improves, the user’s behavior may become more careless. This undermines the entire purpose of using AI for informed decision-making. For the broader AI ecosystem, this finding challenges the narrative that “better models” alone solve trust issues. It points to a user-side failure mode that must be addressed through design, not just training data.

Implications for AI Practitioners

For developers and product managers building health-focused AI tools, this study provides a clear warning: do not optimize solely for answer accuracy. You must also design for user vigilance. Several practical takeaways emerge:

• Design for Skepticism, Not Just Convenience. Interfaces should actively encourage verification. This could mean prominently citing sources, offering disclaimers that are not easily ignored, or even introducing friction (e.g., “Are you sure you want to act on this advice?”) for high-risk queries.
• Measure Dependency, Not Just Satisfaction. Standard UX metrics like “time on task” or “user satisfaction” may actually correlate with over-trust. Practitioners should add metrics that track whether users are critically engaging with the output—such as click-through rates on cited sources or follow-up questions that challenge the AI.
• Consider Adaptive Trust Calibration. Future systems could dynamically adjust their tone or level of certainty based on the user’s apparent dependency. If a user asks the same question repeatedly without verifying, the AI could shift from a confident answer to a more cautious, educational response.

Key Takeaways

• Learned dependency is a real risk: Repeated use of GenAI for health info can degrade a user’s ability to critically evaluate outputs, leading to over-trust.
• Accuracy is not enough: Even a highly accurate model can be dangerous if it trains users to stop thinking critically.
• Design for vigilance: AI interfaces must actively encourage verification and skepticism, especially in high-stakes domains like health.
• New metrics needed: Practitioners should measure user dependency and critical engagement, not just satisfaction or speed.