📄 medicine

When Do Physicians Trust AI? The Role of Clinical Context in Decision Support: SMART vignette study

This study of 420 U.S. physicians reveals that prior exposure to AI, rather than the specific clinical context or amount of information provided, is the primary factor positively influencing physicians' confidence in and trust of AI-enabled clinical decision support tools.

Original authors: Bohye Kim, Katie Ryan, Max Kasun, Jane Paik Kim

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bohye Kim, Katie Ryan, Max Kasun, Jane Paik Kim

Original paper licensed under CC BY 4.0 (https://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

In modern medicine, doctors are beginning to use a new kind of assistant: artificial intelligence. These computer programs, often called clinical decision support systems, can scan patient records and suggest next steps, such as whether to prescribe a medication or refer a patient to a specialist. The promise is that these tools will help doctors make better, faster choices. However, a tool is only useful if the person using it trusts it enough to listen. If a doctor ignores the computer's advice because they do not understand it or do not believe it is right, the technology fails. The big question for hospitals and researchers is not just whether the software works on a computer, but what makes a human doctor feel confident enough to use it in the real world. Is it the type of medical problem? Is it how much information the doctor gets about how the computer thinks? Or is it simply whether the doctor has used this kind of technology before?

A team of researchers at Stanford University set out to answer these questions by asking four hundred and twenty family and internal medicine doctors in the United States to imagine themselves in specific medical situations. They did not ask the doctors to use a real computer program on real patients. Instead, they used a carefully designed survey method where doctors read short stories, or vignettes, about a thirty-three-year-old woman named Jess who was showing symptoms of a health issue. In some stories, Jess needed a prescription for depression, while in others, she needed a referral to a specialist for attention issues. The researchers changed the details of the story to test three different things. First, they changed the type of medical decision, comparing a prescription to a referral. Second, they changed how much information the doctor received about the artificial intelligence tool itself; some doctors got a short description, while others got a much longer one that included details about where the tool was built and how it was trained. Third, they looked at the doctors' own history, noting which ones had used artificial intelligence in their research or practice before.

The researchers wanted to see if these changes would make the doctors feel more or less confident in their final decision. They hypothesized that doctors would feel more confident if they knew more about how the tool worked, and that they would be more skeptical of the tool when the decision involved higher risks, like prescribing a new drug. They also wondered if doctors who were already familiar with artificial intelligence would react differently to the amount of information provided. The study found that the amount of information did not matter as much as expected. Whether a doctor read five sentences or ten sentences about the tool, their confidence in the decision remained the same. The type of medical decision also did not change how confident the doctors felt overall, but it did change how they viewed the tool itself. When the scenario involved prescribing medication, doctors were less likely to trust that the tool would give consistent advice for every patient and were more likely to rely on their own judgment rather than the computer's suggestion. In contrast, when the scenario involved a referral, they were more willing to trust the tool's consistency.

The most significant finding, however, came from looking at the doctors' past experiences. The study showed that doctors who had already used artificial intelligence in their research or clinical practice felt significantly more confident in their final decisions when using the tool. These experienced doctors were also more likely to say that the tool improved the care they could provide and that the computer's result contributed to their final choice. This suggests that familiarity breeds confidence. Interestingly, the researchers found that prior experience did not make doctors trust the tool more blindly; rather, it made them feel more secure in their own ability to use the tool effectively. The study also ruled out the idea that simply dumping more technical details about the software onto a doctor would make them trust it more. In fact, the extra information had no measurable impact on their confidence or their perception of the tool's effectiveness.

These results suggest that the path to integrating artificial intelligence into medicine is less about perfecting the user manual and more about preparing the people who will use it. The researchers conclude that training future doctors should focus on how to weigh computer suggestions alongside their own clinical judgment, rather than just teaching them how to operate the software. If doctors are to trust these tools, they need to understand how to use them in context, and their confidence seems to grow naturally with experience. The study highlights that while the technology is advancing rapidly, the human element remains the critical factor. For artificial intelligence to truly help patients, the medical community must ensure that doctors feel ready and willing to work alongside it, a readiness that appears to come from exposure and practice rather than just more data.

Technical Summary: When Do Physicians Trust AI? The Role of Clinical Context in Decision Support

Problem Statement
Despite the increasing integration of Artificial Intelligence-enabled Clinical Decision Support Systems (AI-CDSS) into healthcare workflows, the translation of algorithmic performance into meaningful clinical impact remains limited, with fewer than 2% of AI models advancing beyond prototyping. A critical barrier to adoption is the "translation gap" driven by human factors: specifically, how clinicians critically evaluate, interpret, and integrate algorithmic outputs with independent clinical judgment. While regulatory approvals surge, successful implementation is heavily shaped by clinician trust, perceived workflow disruptions, and the risk of "alert fatigue," where physicians override alerts due to cognitive burden rather than clinical reasoning. Furthermore, existing literature often examines factors influencing physician trust (such as risk perception, information framing, and prior exposure) in isolation, failing to account for their interrelationships or interactions within real-world clinical contexts. This study addresses the urgent need to understand how clinical decision type (risk profile), the amount of information disclosed about the AI, and prior physician exposure to AI collectively influence confidence, perceived effectiveness, and reliance on AI tools.

Methodology
The study employed a Sequential Multiple-Assignment Randomized Trial (SMART) vignette design, a novel survey methodology adapted for online research.

Participants: The study surveyed 420 active U.S. physicians specializing in Family Medicine or Internal Medicine, recruited via the American Medical Association (AMA) Physician Professional Data file.
Experimental Design: Participants were presented with a series of sequential micro-scenarios ("subvignettes") rather than a single static narrative. Through probabilistic assignment, the study systematically varied three key characteristics across the sample:
1. Clinical Decision Type: Prescription (higher risk) vs. Referral (lower risk).
2. Amount of Information: "More" information (10 sentences, including development data, training data, and AUC values) vs. "Less" information (5 sentences, focusing on usage and patient records).
3. AI Recommendation: Treat vs. Do Not Treat.
Measures: Primary outcomes included self-reported confidence in the final decision and perceived effectiveness of the AI-CDSS (using 5-point Likert scales). Exploratory outcomes assessed perceptions of AI use in practice, reliance on own judgment vs. AI, and willingness to use the tool.
Predictors: The analysis focused on clinical decision type, information volume, and a binary classification of prior AI exposure (defined by self-reported use in research or practice, or high familiarity scores).
Statistical Analysis: Linear regression models were used to assess main effects and interaction terms (moderation). One-sided tests with Benjamini-Hochberg correction were applied for primary hypotheses, while two-sided tests were used for exploratory analyses.

Key Results

Impact of Prior Exposure: Physicians with prior exposure to AI demonstrated significantly higher confidence in their final decisions ( $\beta = 0.274$ , adjusted $p = 0.042$ ) and rated the AI-CDSS as more effective in improving patient care ( $\beta = 0.350$ , adjusted $p = 0.006$ ). Prior exposure was also associated with a greater likelihood of agreeing that the AI result contributed to the final decision ( $\beta = 0.305$ , $p = 0.010$ ).
Impact of Information Volume: Contrary to the hypothesis that more transparency would increase trust, the amount of information provided about the AI-CDSS (more vs. less) had no significant effect on physicians' confidence or perceptions of effectiveness.
Impact of Clinical Decision Type: The type of clinical decision significantly influenced trust and reliance. Physicians assigned to the prescription scenario (higher risk) reported:
- Lower trust in the AI's ability to provide consistent recommendations ( $\beta = -0.211$ , $p = 0.027$ ).
- Fewer perceived benefits outweighing risks ( $\beta = -0.179$ , $p = 0.042$ ).
- Greater reliance on their own examination versus the AI recommendation ( $\beta = 0.154$ , $p = 0.027$ ).
- In contrast, those in the referral scenario showed higher trust and lower reliance on independent judgment.
Moderation Effects: Prior AI exposure did not significantly moderate the effect of information volume on confidence or effectiveness ratings.

Significance and Claims
The paper posits that the successful deployment of AI-CDSS requires a pre-deployment assessment of human factors, specifically highlighting that prior knowledge and exposure to AI are critical drivers of physician confidence and perceived tool effectiveness. The findings suggest that:

Training is Essential: Integrating AI use and AI-assisted decision-making into physician education is vital. The authors argue that training should focus on weighing AI-generated input alongside patient context and clinical judgment, rather than deferring to algorithms.
Context Matters: Physicians differentiate between risk profiles; they are more skeptical and less reliant on AI in high-stakes scenarios (e.g., prescribing) compared to lower-stakes scenarios (e.g., referrals). Implementation strategies must account for these task-specific variations in trust.
Information Framing Limits: Simply providing more technical details about an AI model (e.g., training data, AUC) does not necessarily enhance trust or perceived effectiveness among busy clinicians, suggesting that other interface or presentation factors may be more influential.
Safety and Calibration: To prevent inappropriate reliance on AI, which poses risks to patients, safeguards must be developed to align physician confidence with a model's actual reliability. This includes audits and feedback mechanisms.

The study concludes that while algorithmic performance is necessary, the "human-in-the-loop" dynamics—specifically the physician's background, exposure, and the specific clinical context—are the determining factors in whether AI tools are adopted, scrutinized, or ignored in clinical practice.

Technical Summary: When Do Physicians Trust AI? The Role of Clinical Context in Decision Support

More like this