Nursing expertise makes AI meaningful.

Fact check date: August 22, 2026. All references verified in PubMed and MEDLINE.

IT’S 3:14 AM, and a small red flag appears at the top of your patient’s chart. A number sits next to it, an early warning score, a sepsis probability, a readmission risk, a falls risk—take your pick—and it’s climbed just high enough to interrupt the care you’re providing to another patient.

I’ve talked with enough nurses to know exactly what happens in that moment. Some walk straight to the bedside, no hesitation. Others glance at the score, glance at the patient through the doorway, and think, “She looked fine 10 minutes ago.” Both instincts are reasonable, and both are, on their own, incomplete. That gap between what the algorithm says and what your hands, eyes, and years of pattern recognition are telling you is what this column is about.

Last month I wrote about AI tools already on your unit. This month we go one layer deeper, into the single most common way AI actually touches your practice: Alerts. The number that fires, uninvited, in the middle of your shift and asks you to respond.

The number that interrupts your shift

According to Edelson and colleagues, every alert attached to a patient—a Modified Early Warning Score, a National Early Warning Score (NEWS2), a sepsis flag, a deterioration index, a falls score—does the same basic job. It takes data (heart rate, respiratory rate, blood pressure, oxygen saturation, skin turgor, temperature, level of consciousness, sometimes lab trends) generated from your monitors and documentation and runs them through an algorithm—a set of rules or a trained mathematical model developed to analyze data and produce an output. In this case, the goal is one number that answers a single question: Is this patient condition showing a measurable trend?

However, the model is only as good as the data it receives. The model has no sense of the family member upset in the corner of the room or the pain your patient won’t name out loud. Its output depends entirely on the accuracy and completeness of the data flowing in from connected devices.

The traditional early warning scores, such as NEWS2, have been in use for over a decade. They’re built on fixed thresholds set by clinical consensus. What’s changed in the past several years is that the growing numbers of these scores are no longer purely rule based. They’re increasingly powered by machine learning models, a type of AI, trained on large volumes of patient data, built to catch subtler, earlier patterns of deterioration than what a threshold-based score can identify.

A randomized evaluation by Edelson and colleagues compared a traditional early warning score with an AI-enhanced version in hospital units. The AI-based score produced fewer false alarms and identified patient deterioration earlier. However, these improvements in the tool’s ability to predict deterioration didn’t lead to equally large improvements in patient outcomes. In other words, better prediction doesn’t necessarily mean better outcomes. That distinction matters, and I’ll come back to that.

Here’s the part I want you to consider. None of these tools, rule-based or AI-based, diagnose anything. They’re not telling you what’s wrong with your patient. They’re telling you, with a certain level of statistical confidence, that something in this patient’s data resembles the pattern seen in other patients who later deteriorated. The data may or may not be a large language model, a type of AI, that has all the necessary data to provide a valid answer. That’s a meaningfully different claim than this patient is septic or this patient is decompensating. It’s a nudge, not a verdict.

The cost of too many alarms

If nurses treated every nudge as urgent, no unit could function. Alert fatigue isn’t new. Monitoring technology has improved but the underlying problem, too many alarms competing for the same limited attention, hasn’t disappeared. If not implemented carefully, AI has the potential to make it dramatically worse.

Ancker and colleagues examined how workload, task complexity, and sheer repetition of alerts affected clinician response to a hospital’s clinical decision–support system. As alert volume and repetition rose, the rate at which clinicians responded to the alerts fell sharply. Despite the relevancy of the alerts, their volume overwhelmed clinicians’ capacity to respond meaningfully. Now you have the cry wolf effect: When alarms sound frequently without indicating a real problem, clinicians may stop responding, potentially missing the alarm that indicates something really is wrong.

Downing and colleagues’ randomized evaluation of an electronic health record sepsis alert found that it increased the rate at which nurses and physicians ordered sepsis-related labs and fluids, but it didn’t significantly change mortality, length of stay, or ICU transfer rates compared with usual care. The authors didn’t interpret the alert as useless, but it clearly changed behavior. They noted that an alert alone, disconnected from a broader clinical workflow and a clinician’s own judgment about the whole patient, can’t sufficiently change outcomes. The alert can get a nurse to the bedside faster, but it can’t substitute for what happens once the nurse has arrived.

This is precisely why some health systems have begun tiering their AI-based alerts, reserving hard interruptions for high-confidence, high-acuity signals and routing lower-confidence signals to a passive dashboard a charge nurse can scan. If you feel your unit’s system cries wolf, bring it up in a shared governance or informatics committee meeting and document it in writing to your leadership. Alert design isn’t fixed in stone, and nurses are exactly the people who should shape it.

The quiet risk of believing the score too much

Alert fatigue can cause clinicians to pay too little attention to alerts. Automation bias can create the opposite problem—placing too much trust in an algorithm or an AI-generated prediction or recommendation, even when a clinician’s assessment suggests otherwise. This is where your judgment matters most because you, not the software and not the institution, are accountable.

Automation bias refers to the tendency to over-rely on automated cues, either by failing to notice when the automation is wrong (omission error) or by following its recommendation even in the face of contradictory information (commission error). An editorial by Khera and colleagues put it bluntly: Assistive AI tools that are wrong with confidence are, in some ways, more dangerous than tools that are wrong with hesitation, because clinicians are trained to trust confident numbers. A risk score reading 72% carries a false sense of precision. It sounds more authoritative than a nurse’s own assessment of a patient’s affect, skin turgor, or work of breathing, and that authority isn’t always earned. In other words, numbers aren’t more valid than nursing documentation.

A second, quieter trap also exists. Hospital policy is sometimes built into an AI model’s rules before your state’s nurse practice act ever enters the conversation. Those two separate documents aren’t interchangeable. Know your practice act first, then know your institution’s policies, especially if you travel or hold licensure in more than one state. A model trained on one system’s rules doesn’t know what your specific state requires of you.

Who’s accountable when the tool is wrong

Contracts between healthcare organizations and technology companies frequently contain a clause stating plainly that the vendor is a technology provider only and that all clinical judgment remains with the licensed clinician. I want to be unambiguous about what that means in practice, because the underlying literature and the American Nurses Association Code of Ethics for Nurses agree on this point: The tool doesn’t hold a license. You do. Following an alert’s recommendation doesn’t transfer your accountability to the software vendor, and neither does overriding one provided your clinical reasoning is sound and documented. (See Technology agreements.)

I want to be direct about why this matters for you and not just for the informatics team that configured the tool. You’re the one physically present. You’re the one who can see that a patient’s tachycardia is anxiety about a family conversation happening down the hall, not sepsis. You’re the one who knows this particular patient always runs a low-grade fever after surgery. A simple algorithm or AI-enhanced alert, trained on population-level patterns, has never met this patient before today. You have. When something doesn’t add up, go back to basics. Take the blood pressure manually before making a decision based only on what the monitor reports. Assess the patient and environment quickly to determine the alert’s validity.
Staffing and surveillance capacity

Let me say it plainly. No early warning score, AI augmented or otherwise, changes who’s accountable for what happens next. That accountability sits with you, the licensed nurse at the bedside exercising independent clinical judgment. A high score doesn’t discharge that responsibility and neither does a reassuring one.

An older, blunter body of literature about the failure to rescue belongs in this conversation. Garvey described it as the type of scenario an early warning score is designed to prevent—a patient who develops a complication that, had it been recognized and acted on promptly, wouldn’t have progressed to serious harm. A classic nursing analysis of failure to rescue makes a point that has never stopped being true. Ongoing surveillance, pattern recognition, and a willingness to escalate concerns, even against resistance, separate a rescued patient from a missed one.

The evidence behind that claim isn’t abstract. A landmark study by Aiken and colleagues of more than 230,000 surgical patients across 168 Pennsylvania hospitals found that for every additional patient added to a nurse’s assignment, the odds of a patient dying within 30 days of admission rose by 7%, and the odds of failure to rescue rose by that same 7%.

These researchers found that increased workload carried a cost for nurses, raising the odds of burnout by 23% and job dissatisfaction by 15%. Mortality risk climbed with each additional patient added to a nurse’s workload. These findings highlight a staffing issue, not a credentialing one. Adequate staffing gives nurses the capacity to monitor patients, recognize changes, and respond to early warning scores. Technology can support that surveillance, but it can’t replace the nurses needed to provide it.

An AI-based alert can’t solve inadequate staffing. At its best, it can help nurses monitor patients and identify changes—but only when nurses have the time and capacity to respond.

What you do next

The alert fires. What does that look like in practice?

1. Look at the patient, not the alert. A score is generated from minutes-old data.
2. Conduct a focused reassessment tied to whatever domain triggered the alert. If it’s respiratory, listen to the lungs and count the respiratory rate rather than trusting the monitor’s derived number. If it’s a sepsis flag, look at the clinical picture—mentation, capillary refill, and skin turgor, temperature, and color—which no single vital sign captures on its own.
3. Put your findings into words a provider or rapid response team can act on immediately. Situation, Background, Assessment, Recommendation (SBAR) remains the most reliable nursing structure. For example, you might say: “The system flagged a NEWS2 of seven. My own assessment adds these findings: [insert findings]. Notifying a provider-physician or nurse practitioner to evaluate now.” That’s stronger and more useful communication than simply forwarding the number.
4. Document your own independent assessment, not just the system-generated score. If you agreed with the alert, document what you personally observed that supported it. If you disagreed and chose watchful waiting, document the specific clinical reasoning behind that choice. Clinical judgment strengthens with each documented encounter over the course of a career. That documentation records how you exercised nursing judgment rather than simply reacting to, or ignoring, a machine. It’s also, frankly, your best protection if the patient’s course requires review later.
5. If your judgment and the algorithm or AI-enhanced alert genuinely disagree, escalate the disagreement itself, not just the patient’s status. Say out loud to the charge nurse, the rapid response team, or the provider, “The score says X, but here’s what I’m seeing.” That sentence isn’t insubordination toward a piece of software. It’s the exact clinical judgment your license and your state’s nurse practice act require.

Carry this into your next shift

A high-tech alert doesn’t lower the bar for nursing judgment. It raises the floor of information you’re working from. The number is a prompt, not a diagnosis. It tells you where to look, not what you’ll find.

Your independent assessment and documentation of that assessment are what the record, and possibly your license, will lean on. Write down what you saw, not just what the system said. This serves as the foundation for whatever comes next in nursing documentation because future AI tools will only be as useful as the data nurses put into them today.

If your clinical assessment and the algorithm or AI-enhanced alert disagree, say so out loud in words a colleague can act on. That disagreement, spoken clearly, isn’t a failure of the technology. It’s the technology working exactly as intended, as a second set of eyes, not a first and only opinion.
The early warning system on your unit didn’t replace the skill that’s always separated a rescued patient from a missed one. It just gave that skill a little more lead time. Use it.

Roy’s Rule: AI may be powerful, but nursing expertise makes it meaningful.

Roy L. Simpson is a nurse informatician with 5 decades of experience at the intersection of nursing and technology. He currently serves as professor, co-director, and assistant dean for AI education at the Nell Hodgson Woodruff School of Nursing at Emory University in Atlanta, Georgia.

American Nurse Journal. 2026; 21(11). Doi: 10.51256/ANJ112629

References

Aiken LH, Clarke SP, Sloane DM, Sochalski J, Silber JH. Hospital nurse staffing and patient mortality, nurse burnout, and job dissatisfaction. JAMA. 2002;288(16):1987-93. doi:10.1001/jama.288.16.1987

American Nurses Association. Code of Ethics for Nurses. 2025. codeofethics.ana.org/home

Ancker JS, Edwards A, Nosal S, et al. Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. BMC Med Inform Decis Mak. 2017;17(36). doi:10.1186/s12911-017-0430-8

Downing NL, Rolnick J, Poole SF, et al. Electronic health record-based clinical decision support alert for severe sepsis: A randomised evaluation. BMJ Qual Saf. 2019;28(9);762-8. doi:10.1136/bmjqs-2018-008765

Edelson DP, Churpek MM, Carey KA, et al. Early warning scores with and without artificial intelligence. JAMA Netw Open. 2024;7(10):e2438986. doi:10.1001/jamanetworkopen.2024.38986

Garvey PK. Failure to rescue: The nurse’s impact. MEDSURG Nurs. 2015;24(3):145-9.

Khera R, Simon MA, Ross JS. Automation bias and assistive AI: Risk of harm from AI-driven clinical decision support. JAMA. 2023;330(23):2255-7. doi:10.1001/jama.2023.22557

Key words: artificial intelligence, AI, alerts, early warning scores