Over time and with experience, physicians learn when to trust algorithms

As organizations invest billions in artificial intelligence, they face a problem that no algorithm can solve for them: convincing people to use it. New research shows that demonstrating an algorithm’s accuracy is not enough. Nonetheless, reluctance to follow its recommendations may diminish over time, especially in situations where professionals retain the ability to exercise judgment. It is in these contexts, in fact, that experience gained from working with AI appears to progressively change the way people integrate the algorithm’s guidance into their own decisions.

The study, conducted by Luigi Preti, Amelia Compagni, and Giulia Cappellaro of CERGAS, SDA Bocconi, and Bart Vanneste of the UCL School of Management, analyzed more than 650,000 recommendations generated by an AI system used at Fresenius Medical Care dialysis centers to support nephrologists’ treatment decisions in the care of dialysis patients. The main finding is that, over time, clinicians become more likely to follow the algorithm’s advice precisely in situations where clinical guidelines leave more ample room for professional judgment. But this doesn’t mean that doctors are becoming more obedient to machines.

At a time when artificial intelligence is entering decision-making processes (and not only in healthcare but also in businesses, banks, insurance companies, and public administrations), the research suggests that the success of AI depends as much on the technology as on organizations’ ability to guide people through implementation and learning.

Many tools, limited use

In recent years, artificial intelligence has rapidly made inroads into clinical practice. In the United States, more than 1,500 AI-based medical software products have been authorized, primarily to support clinicians in diagnostic and prognostic decisions. Yet many of these tools continue to be underused in everyday practice, limiting their potential impact on the quality of care.

The literature has long discussed algorithm aversion, meaning people’s reluctance to rely on algorithmic guidance even when algorithms demonstrate strong performance. The phenomenon has been observed in many fields, from finance to the justice system, but it is particularly relevant in healthcare, where decisions directly affect patients’ health.

Most previous research has focused on algorithms that help make diagnoses or predictions, for example in the analysis of radiological images. But two questions were left largely unexplored:

• What happens when artificial intelligence comes into play in treatment decisions where physicians retain considerable room for judgment?
• Does the relationship between experience with an algorithm and willingness to follow its recommendations change over time?

AI recommendations and their effects

The researchers studied a clinical decision support system used to manage anemia in dialysis patients. Each month, based on the patient’s blood tests and clinical history, the algorithm suggests to the nephrologist the dosage of a specific drug to administer. Compared with the previous month, the AI guidance may involve maintaining, increasing, or reducing treatment. The physician is free to accept or reject the advice, but if they decide not to follow it, they are asked to provide a reason.

The study is based on an exceptionally large dataset: 659,976 recommendations concerning 27,819 patients, made by 1,061 nephrologists working at 171 dialysis centers in eight European countries, observed between 2013 and 2024.

The authors distinguish between the simple number of recommendations and the amount of time spent working with the algorithm. This distinction is pertinent because, in treatment monitored over time such as anemia management, physicians do not merely get AI advice; they also have the opportunity to observe in subsequent months how the patient’s condition evolves.

Gaining experience while continuing to make the decisions

The main finding shows that greater experience with the system over time is associated with a higher probability of accepting its recommendations. In cases when the time elapsed since the system was first used is doubled, this is associated with an average increase of about 1.5 percentage points in the probability that a physician will accept the algorithm’s proposal. The effect is gradual and far from widespread: In the study sample, the average acceptance rate is 56.5%, confirming that medical professionals still frequently ignore the system’s advice.

Even more interesting is the circumstances in which this experience/acceptance relationship emerges. To explain, when clinical guidelines strongly prescribe which decision should be made, the amount of time spent with the algorithm is essentially unrelated to the probability that clinicians will accept its recommendations. The positive relationship arises instead in situations where the guidelines do not unequivocally dictate the action to take and the physician has greater discretion. It is precisely in these contexts that experience appears to progressively change the relationship between professional judgment and algorithmic advice.

By looking at the reasons clinicians give for rejecting AI output, we can get a better understanding of the phenomenon. When lab results are within the desired range, the most common reason for not following the algorithm is that the patient is stable and, in the physician’s opinion, it isn’t appropriate to change a treatment that’s working. When anemia needs to be reduced, by contrast, physicians often agree with the direction of the change proposed by the algorithm but believe that the dosage adjustment should be more gradual.

In any case, experience is not simply associated with a uniform reduction in the number of rejections. It also changes the way clinicians depart from the algorithmic recommendations: Over time, they continue to exercise their own judgment and distinguish between different clinical circumstances. In other words, the findings do not suggest that physicians’ judgment is replaced by the machine’s, but rather that the way the two are integrated may change over time.

The study also reveals that this evolution is accompanied by a greater probability that patients’ hemoglobin levels will fall within the recommended range. If a physician had a greater previous propensity to accept the algorithm’s advice, this also corresponds to a higher probability of reaching the target. Because these are observational data, the findings do not demonstrate that greater acceptance of AI recommendations causes better clinical outcomes, but the results do show that the observed evolution in AI use is consistent with improved anemia management.

Collaboration between professionals and AI

The success of an artificial intelligence project is most often attributed to the quality of the algorithm. This research suggests, however, that the way people learn to use it also matters.

AI adoption is a process. To successfully introduce an AI system, it may be important to create the conditions for people to work with it long enough to progressively understand its strengths and limitations and observe how decisions translate into outcomes. This means that beyond initial training, also essential are continuity, feedback, and opportunities to gain experience.

The study also suggests that these strategies should not necessarily be identical for all types of decisions. Experience appears to matter most when professionals still have leeway to exercise their own judgment, and they must decide how to factor in the algorithmic suggestions with their own knowledge of the case. When guidelines already clearly prescribe the action to take, by contrast, there is much less scope for experience with AI to change behavior.

While the debate over artificial intelligence has so far focused primarily on whether algorithms can make better decisions than humans, this study shifts attention to how collaboration between professionals and algorithms evolves over time. Experience can help professionals integrate artificial intelligence guidance more effectively into their own judgment, without giving up their ability to depart from these recommendations when clinical circumstances warrant it.

The study When Does Experience Attenuate Algorithm Aversion? Evidence from AI-supported Therapeutic Decision Making in Nephrology Care, by Luigi Preti, Amelia Compagni, Giulia Cappellaro, and Bart Vanneste, received the Best Paper Award from the Healthcare Management Division at the Academy of Management Annual Meeting 2026, which has just concluded in Philadelphia, as well as the Best Paper award at the Digital Transformation Society Conference organized by Paris Business School and the University of Naples Parthenope.

Luigi Preti, Amelia Compagni, Giulia Cappellaro, and Bart Vanneste. “Experience With AI Algorithms And Algorithm Aversion In AI-Supported Medical Care.” Academy of Management Proceedings. DOI: https://doi.org/10.5465/amproc.2026.274bp.