AI Alerts Versus Clinical Judgment: A Point-Counterpoint on the New Medical Error

When the Algorithm Speaks First

Decision often begins before the clinical encounter starts. I open the chart and immediately see a banner alert identifying a patient as high risk for sepsis or clinical deterioration. Before I have reviewed the history or examined the patient, the system has already offered an interpretation. Without realizing it, I began to internalize that interpretation as well.

For much of my career, I viewed clinical decision support tools as an unquestioned benefit to patient care. They promised earlier detection of illness, more standardized management, and fewer missed diagnoses.^1,2 In a specialty defined by cognitive complexity and constant interruption, the idea of an ever-present digital assistant felt not only helpful but necessary.

Over time, I began to notice a more subtle effect. These alerts were not simply informing my decisions. They were shaping them. What once felt like background support gradually became a foreground influence. The presence of an alert altered the starting point of clinical reasoning. Instead of asking what might be happening with this patient, I found myself asking why the alert had been triggered.

This shift may seem minor, but it represents a meaningful change in how we think.

Clinical reasoning traditionally begins with the patient and builds outward through history, examination, and synthesis. Decision support systems begin with a conclusion and ask us to confirm or reject it.^1,3 When these two processes intersect, the algorithm often sets the frame for the encounter.

A new type of medical error begins to emerge. It is not an error of knowledge or negligence. It is an error of deference. Clinicians do not blindly follow alerts, but the presence of an alert changes the threshold for action.³ It introduces a subtle pressure to respond, to intervene, and to justify inaction if we choose not to. I have observed antibiotics initiated not because the clinical picture clearly supported infection, but because an alert suggested the possibility.^4,5 I have seen additional tests ordered to investigate a risk signal that did not align with bedside assessment. I have also seen hesitation, not due to uncertainty, but due to discomfort with overriding a system recommendation.³

Each of these decisions is individually defensible. Collectively, they reflect a gradual shift in clinical behavior. We begin to practice in response to signals rather than in response to patients.

What makes this phenomenon difficult to address is that it rarely appears as a traditional error. There is no guideline violated. The documentation remains appropriate. Care decisions can be justified. Yet, the trajectory of care changes in ways that are subtle but consequential. Overtreatment, diagnostic cascades, and resource utilization increase without a clear moment when something has gone wrong.

Decision support systems are designed to promote consistency.^2,5 However, consistency is not always synonymous with appropriateness. Patients do not present as averages. They present as individuals with unique physiology, context, and variability that often fall outside predictive models.^1,3

The presence of an alert conveys a sense of certainty, even when the underlying signal is problematic. Over time, this can shift the clinician’s role from independent decision maker to interpreter of algorithmic output.³ That shift is not trivial. Supervision of a system is fundamentally different from ownership of a clinical decision.

The concern is not that artificial intelligence will replace physicians. The more immediate concern is that it will reshape clinical reasoning in ways that are difficult to recognize in real time.^1,3 When attention is directed by alerts rather than by independent assessment, the discipline of clinical judgment begins to erode.

This is how a new category of medical error takes hold. It does not arise from a lack of knowledge. It arises from an overreliance on tools designed to assist us.³ The irony is that the very systems intended to improve care can, under certain conditions, subtly distort it. 

Dr. Patel

Dr. Patel is a hospitalist at Springfield Clinic in Springfield, Ill., and president of the SHM Chicago chapter.

In Defense of the Algorithmic Assistant

The concerns raised by my colleague are thoughtful and deserve careful consideration. Clinical decision support systems do influence behavior, and there are circumstances in which they contribute to unnecessary interventions.³ However, focusing only on these risks overlooks the broader context in which these tools have been developed.

Modern hospital medicine places extraordinary cognitive demands on clinicians. Patients are more complex, data streams are more abundant, and the pace of care is relentless. Clinicians are expected to synthesize vast amounts of information while managing interruptions, documentation requirements, and competing priorities.¹ In this environment, unaided clinical reasoning is not always sufficient.

Clinical decision support should not be viewed as a replacement for judgment, but as a response to the limitations of human cognition under strain.¹ Even experienced clinicians are vulnerable to cognitive biases such as anchoring, availability, and premature closure. These are not rare occurrences. They are inherent features of decision making in high-pressure environments.

When used appropriately, decision support systems can serve as a counterbalance to these limitations. An alert may prompt reconsideration of a diagnosis that might otherwise have been overlooked. A risk score may highlight patterns that are not immediately apparent during a busy shift.^6,7 These tools do not eliminate error, but they can shift the likelihood of certain types of error occurring.

Equally important is the role of consistency. Before the widespread use of decision support, recognition of clinical deterioration often depended on individual vigilance. This introduced variability in care that could affect patient outcomes.^4,6,7 Decision support systems help ensure that certain signals are consistently identified and acted upon across providers and settings.

They also introduce a degree of accountability that was previously difficult to achieve. Alerts create a record of when a potential concern was raised. This visibility allows for retrospective learning and system improvement.³ It transforms what was once implicit into something observable and actionable.

Rather than viewing current challenges as evidence of failure, it may be more accurate to see them as evidence of transition. We are in the early stages of integrating these tools into clinical practice. The difficulties we encounter reflect not a fundamental flaw in the concept, but an incomplete adaptation to its use.

Every major technological advancement in medicine has required a period of adjustment. Imaging technologies introduced challenges related to overdiagnosis. Laboratory testing expanded the problem of false positives. Over time, clinicians developed frameworks to interpret and apply these tools more effectively.¹ Clinical decision support is following a similar trajectory. The goal is not blind adherence, nor is it wholesale rejection. The goal is thoughtful integration.

This requires the development of new competencies. Clinicians must understand not only what an alert indicates, but how it is generated and what its limitations are.³ Systems must be designed with attention to specificity and usability.³ Education must evolve to include the interpretation of algorithmic outputs alongside traditional clinical training.^1,3

When these elements are aligned, decision support systems can enhance rather than diminish clinical judgment. They can extend our capacity to detect risk, support decision making under pressure, and reduce variability in care.^2,4,5,7

New forms of error may indeed emerge. That is an inevitable consequence of any new tool. However, it is equally true that opportunities to improve care will expand.

The central question is not whether these systems influence clinical reasoning. They do.³ The more important question is whether we are willing to develop the skills and systems necessary to use them effectively.

Clinical judgment has never been static. It evolves with evidence, experience, and technology. The integration of decision support represents not a loss of judgment, but a redefinition of how it is applied.¹

The Bottom Line

Clinical decision support is reshaping the landscape of hospital medicine in ways that are both promising and challenging.^1-3 One perspective highlights the risk of overreliance and the subtle erosion of independent thinking. The other emphasizes the realities of cognitive limitation and the need for tools that support clinicians in increasingly complex environments.

These perspectives are not mutually exclusive. They reflect two sides of the same transformation. The future of hospital medicine will not be determined by whether we adopt these tools, but by how we use them.^1,3 The essential task is to ensure that technology informs clinical reasoning without replacing it.

Clinical judgment must remain the foundation of decision making. Decision support should guide attention, not dictate action.³ The distinction is critical.

In the end, the goal is not to practice medicine with or without algorithms. It is to practice medicine in a way where responsibility remains clearly human, even as the tools we use become increasingly sophisticated. That responsibility cannot be delegated. It must be retained, deliberately and consistently, at the bedside. 

Dr. Brodkin

Dr. Brodkin is an ambassador flex medical director at Sound Physicians, currently working in London, Ky., as a traveler. He is also the vice president of the SHM Chicago chapter.

References

1. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56. doi:10.1038/s41591-018-0300-7.
2. Shahmoradi L. Clinical decision support systems-based interventions to improve medication outcomes: A systematic literature review on features and effects. Medical Journal of The Islamic Republic of Iran. 2021; doi:10.47176/MJIRI.35.27
3. Edelson DP, et al. Early warning scores with and without artificial intelligence. JAMA Netw Open. 2024;7(10):e2438986. doi:10.1001/jamanetworkopen.2024.38986.
4. Amland RC, Hahn-Cover KE. Clinical decision support for early recognition of sepsis. Am J Med Qual. 2019t;34(5):494-501. doi:10.1177/1062860619873225.
5. Bignami EG, et al. Artificial intelligence in sepsis management: an overview for clinicians. J Clin Med. 2025;14(1):286. doi:10.3390/jcm14010286.
6. Tun HM, et al. Trust in artificial intelligence-based clinical decision support systems among health care workers: systematic review. J Med Internet Res. 2025;27:e69678. doi:10.2196/69678.
7. Verma AA, et al. Clinical evaluation of a machine learning-based early warning system for patient deterioration. CMAJ. 2024;196(30):E1027-E1037. doi:10.1503/cmaj.240132.