Andrii Bidochko

• Updated: July 1, 2026
• 6 min read

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

Direct Answer

The paper “Trust in Generative AI for Health Information Consumption and the Effect of Learned Dependency: An Experimental Study” reveals that users who develop a habit of relying on generative AI for medical queries tend to over‑trust the system, even when the output is wrong. The researchers demonstrate that this learned dependency weakens trust calibration and makes people more vulnerable to misinformation, while visual attention cues (e.g., highlighting uncertain passages) do not significantly mitigate the effect.

Background: Why This Problem Is Hard

Generative AI models such as ChatGPT, Claude, and Gemini have become go‑to sources for quick health advice. Their conversational style lowers the barrier to entry, but the same fluency can mask factual errors. In clinical contexts, a misplaced recommendation can lead to delayed treatment, unnecessary medication, or harmful self‑diagnosis.

Existing safeguards—model‑level fact‑checking, post‑generation verification, or simple confidence scores—assume that users will adjust their trust based on the signal presented. Empirical evidence, however, shows that people often ignore or misinterpret these cues, especially after repeated positive experiences. The core difficulty lies in the psychological feedback loop: successful interactions reinforce reliance, which in turn reduces critical scrutiny, creating a “trust drift” that is hard to reverse.

What the Researchers Propose

The authors introduce a two‑factor experimental framework that isolates (1) the accuracy of AI‑generated health content and (2) the presence of visual attention cues designed to draw users’ focus to potential uncertainty. By randomizing 338 participants across a 2 × 2 matrix, they measure two constructs:

• Trust level – captured through a validated Likert‑scale questionnaire after each interaction.
• Learned dependency – quantified by a pre‑study exposure task where participants repeatedly used a health‑focused generative assistant, establishing a habit of reliance.

The framework treats the AI system as a “trust agent” whose output quality and UI signals can be independently manipulated, allowing the researchers to observe how dependency modulates trust calibration.

How It Works in Practice

The experimental workflow can be broken down into three conceptual stages:

1. Dependency induction: Participants complete a series of health‑information queries using a generative model. Correct answers are deliberately mixed with subtle inaccuracies, but the majority are accurate, fostering a sense of reliability.
2. Test phase: In the main study, each participant receives either a correct or incorrect answer to a new health question. Simultaneously, the UI either highlights the answer with a visual cue (e.g., a colored border) or presents it plainly.
3. Assessment: After reading the answer, participants rate their trust in the system and indicate whether they would act on the advice.

This design isolates the psychological impact of “learned dependency” from the immediate effect of answer accuracy. The visual attention cue is intended to act as a low‑cost, scalable mitigation strategy—if users notice a cue, they might pause and verify the information.

What sets this approach apart from prior work is the explicit measurement of dependency as a variable, rather than treating it as a background characteristic. By operationalizing dependency, the study quantifies how habit formation directly skews trust judgments.

Evaluation & Results

The authors applied linear regression models to test four hypotheses:

• H1: Accurate AI output increases user trust.
• H2: Higher learned dependency correlates with higher trust.
• H3: Dependency moderates the relationship between accuracy and trust (i.e., trust calibration).
• H4: Visual attention cues reduce over‑trust in incorrect answers.

Key findings:

• Accurate answers raised trust scores by a statistically significant margin (p < 0.01).
• Participants with stronger dependency scores reported higher overall trust, regardless of answer correctness.
• The interaction term between accuracy and dependency was significant, indicating that highly dependent users exhibited weaker trust calibration—they remained relatively trusting even when presented with false information.
• Visual attention cues did not produce a measurable change in trust levels nor did they moderate the dependency‑accuracy interaction.

In plain language, the experiment confirms that habit‑based reliance erodes the natural “stop‑and‑verify” instinct, and simple UI highlights are insufficient to restore balanced skepticism.

Why This Matters for AI Systems and Agents

For developers of health‑focused conversational agents, the study delivers three actionable insights:

1. Design for calibrated trust: Systems should embed mechanisms that periodically disrupt dependency loops—e.g., random prompts asking users to confirm sources or to seek a second opinion.
2. Beyond visual cues: Since highlighting uncertainty failed to curb over‑trust, richer interventions (explanations, provenance metadata, or interactive fact‑checking tools) may be required.
3. Monitoring dependency metrics: By tracking usage patterns (frequency, repeat queries on similar topics), platforms can infer when a user is entering a high‑dependency state and adapt the interaction accordingly.

These principles align with emerging best practices for trustworthy AI, especially in regulated domains like healthcare. Integrating them into an UBOS platform overview can help enterprises build agents that not only answer questions but also maintain a healthy level of user skepticism.

What Comes Next

While the study sheds light on the psychological underpinnings of AI trust, several limitations remain:

• Sample diversity: Participants were recruited online and may not represent older adults or clinicians who interact with health AI differently.
• Scope of content: The health queries focused on general wellness; high‑stakes scenarios (e.g., medication dosing) could amplify trust drift.
• Intervention depth: Visual cues are a low‑effort solution; future work should explore multimodal explanations, confidence intervals, or third‑party verification APIs.

Future research directions include:

• Longitudinal studies that track dependency over months rather than a single session.
• Cross‑cultural experiments to see how cultural attitudes toward authority affect trust calibration.
• Integration of real‑time fact‑checking services (e.g., leveraging Chroma DB integration) to provide evidence‑backed answers.

Practitioners looking to prototype these ideas can start with the Workflow automation studio to orchestrate dependency‑aware prompts, or explore the Enterprise AI platform by UBOS for scalable deployment across patient‑facing applications.

References

• Ahmed, A., Leroy, G., Sachdeva, A., Harber, P., Rains, S. A., Youn, S., & Barai, P. (2026). Trust in Generative AI for Health Information Consumption and the Effect of Learned Dependency: An Experimental Study. arXiv preprint arXiv:2606.20605.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.