Incident Overview
On February 17, 2026, an AI-generated sepsis alert at an unnamed U.S. hospital triggered a protocol-driven order for intravenous (IV) fluid administration to a dialysis patient. The patient, who was receiving dialysis, was at high risk of fluid overload—a condition that can lead to severe complications such as pulmonary edema or cardiac strain. A clinician, specifically a nurse named Adam Hart, recognized the potential harm and intervened before the fluid bolus was administered. The incident was reported as a near-miss, highlighting how even accurate AI alerts can lead to inappropriate clinical actions if not properly contextualized by human judgment.
The incident occurred at St. Rose Dominican Hospital in Henderson, Nevada, though the exact date remains uncertain, as the event was described as having taken place 'a few years' prior to its public reporting in 2026. The sepsis alert was generated by an unspecified AI system deployed within the hospital’s clinical decision support infrastructure. The system’s recommendation followed standard sepsis protocols, which often include rapid IV fluid resuscitation to stabilize patients. However, the patient’s dialysis-dependent status made such an intervention dangerous, as their kidneys were already impaired and unable to process excess fluids.
The near-miss was averted due to the timely intervention of a nurse and a physician, who overrode the AI’s recommendation. The nurse, Adam Hart, who had 14 years of experience at the hospital, objected to the fluid bolus order, citing the patient’s dialysis catheter and the risk of fluid overload. A physician subsequently ordered an alternative treatment, preventing potential harm. This incident underscores the critical role of clinical expertise in mitigating risks posed by AI-driven decision support systems, even when the AI’s output appears correct.
The case was reported by media outlets, including Scientific American, and later documented in the AI Incident Database (AIID) as Incident 1374. The database categorizes this incident under AI system safety, failures, and limitations, emphasizing how protocol-driven alerts can inadvertently lead to harm when clinical context is overlooked.
Possible Root Cause
The primary root cause of this near-miss appears to be the AI system’s reliance on generalized sepsis protocols without adequate integration of patient-specific contraindications. Sepsis prediction models, including those powered by AI, are typically trained to identify systemic inflammatory responses and trigger interventions like fluid resuscitation based on physiological markers such as heart rate, blood pressure, and lactate levels. However, these models often lack granular clinical context, such as a patient’s dialysis status, which can drastically alter treatment appropriateness.
In this case, the AI system likely followed a standard sepsis protocol that prioritizes fluid administration to restore blood pressure and organ perfusion. Such protocols are designed for patients who can tolerate fluids, but they do not account for scenarios where fluid overload would be catastrophic. The dialysis patient’s impaired renal function meant that their body could not excrete excess fluids, increasing the risk of pulmonary edema or hypertension. The AI’s failure to incorporate this critical information into its recommendation suggests a systemic limitation in how clinical decision support tools interpret patient-specific risks.
AI-generated sepsis alerts often trigger automated or semi-automated orders within electronic health record (EHR) systems, reducing the time clinicians have to assess whether the recommendation aligns with the patient’s overall condition. In this instance, the urgency of the sepsis alert may have overridden the need for a more deliberate evaluation of the patient’s dialysis-dependent status. This highlights a broader issue in AI-driven clinical workflows: the tension between speed of intervention and the necessity of human oversight to prevent protocol-driven errors.
Without such integration, the AI’s recommendation remained disconnected from the patient’s broader clinical picture.
Takeaways
This incident provides several critical insights into the risks and challenges of deploying AI in clinical decision support, particularly in high-stakes scenarios like sepsis management.
Sepsis prediction models, while effective at identifying systemic distress, often operate in isolation from other clinical pathways.
Second, the incident underscores the irreplaceable role of clinical judgment in AI-assisted workflows. Even when AI alerts are accurate, they must be interpreted within the broader framework of a patient’s condition. Hospitals should implement safeguards such as mandatory clinician review for high-risk interventions, particularly in cases where AI recommendations conflict with known patient limitations. Training programs should also emphasize the importance of questioning automated suggestions, especially when they deviate from established clinical guidelines for specific patient populations.
Third, the case highlights the need for better alignment between AI-driven protocols and institutional clinical pathways. Many hospitals rely on standardized sepsis bundles that prioritize fluid resuscitation, but these protocols do not account for variations in patient tolerance.
While this case was averted, similar errors may occur elsewhere if not documented and analyzed.
Finally, this near-miss illustrates the broader ethical and safety implications of AI in medicine. As these systems become more integrated into clinical workflows, their limitations must be clearly communicated to healthcare providers. Clinicians should be trained not only to use AI tools but also to recognize their boundaries—particularly in scenarios where patient-specific factors override generalized recommendations. The goal is not to eliminate AI assistance but to ensure it augments, rather than replaces, human decision-making in ways that prioritize patient safety.
Sources
Incident summary from the AI Incident Database, used under CC BY-SA 4.0.