From technology adoption to noise reduction: Behavioral and psychological dimensions of clinical artificial intelligence deployment.

For more than five decades, patients with the same condition have received markedly different care depending on which clinician they happen to see. This unwarranted clinical variation has been documented in nearly every field of medicine, and decades of standardization efforts have not eliminated it. Artificial intelligence (AI) seemed to promise a fix: an algorithm would give the same answer every time, and randomized trials showed that, on average, AI assistance improved clinician performance. Yet clinicians use these tools far less than expected. The main barrier is now recognized as behavioral, not technical: different clinicians respond to the same AI recommendation in very different ways, and the algorithm's performance does little to predict who will follow it. This narrative review synthesizes the evidence on this obstacle through the perspective of human behavior and attitudes in decision-making. We show that the current challenges of AI deployment are essentially the same challenge of variation in clinical decision-making, now carried over into variation in how clinicians use AI. This produces two opposing errors: accepting an AI recommendation when it is wrong and ignoring it when it is correct. We then describe how these behavioral signals can be fed back into the AI deployment architecture to guide clinicians toward more appropriate use. Every day, across thousands of hospital wards and outpatient clinics, a quiet battle is fought between clinicians and the AI systems deployed to assist them. It is not a battle anyone declared, and neither side fully understands the other. Clinicians of varying age, experience, and psychological disposition encounter algorithmic recommendations with varying feelings, resulting in accepting them, overriding them, or simply ignoring them. The algorithms, on their part, maintain the provision of alerts and recommendations, oblivious to the wishes and needs of the clinicians they serve. The casualties are the patients, who expect consistent, evidence-based care but instead receive whatever combination of random human-algorithm interaction happened to prevail on that particular day, with that particular physician.