Everyone Except the Clinicians
A blanket alert can close a liability gap overnight. What it opens — alert fatigue, and the clinical risk that follows — depends on who was in the room when it was built.
A team launched a healthcare product. Eventually, a set of criteria slipped through, an alert didn’t trigger, and a patient passed away. A lawsuit followed. Like many organizations under pressure, the company tries to close the gap. Every time those criteria appear, an alert popped up—no exceptions, no thresholds, no room for judgment. The fix was built by legal, product, and marketing teams, but no one who would actually be at the bedside when the alert rang.
It went live. Within days, doctors were overwhelmed. Because the alert couldn’t tell what was truly important from what was just noise, it treated everything the same—and clinicians did what anyone would do under alert overload: they started tuning it out. The rollback happened quickly. The solution had created a new kind of clinical risk while trying to fix the old one.
This is a true story, slightly adjusted, shared with me by a healthcare UX researcher who wasn’t directly involved. What made it stick wasn’t just the details—it’s that I’ve heard a version of it from almost everyone who’s worked inside health systems. It’s not just a warning; it’s a well-known pattern with a name and plenty of research behind it.
The pattern has a name
Alert fatigue is one of the most studied issues in clinical informatics, and the findings have stayed pretty consistent over two decades. A review of computerized physician order entry systems showed that clinicians override clinical decision support alerts anywhere from 46% to 96% of the time, with the appropriateness of those overrides varying just as much by alert type.¹ Ancker and colleagues’ classic study on alert responses in outpatient care found two main reasons for the drop-off: mental overload from heavy workloads and complexity, and simple desensitization from repeated alerts that don’t end up mattering.² Both point to the same idea—the more unfiltered signals you send, the less people will use them, whether they’re helpful or not.
This isn’t a settled issue. A 2026 review of alert fatigue showed that even after two decades of research, only one paper offered a clear definition of the term.³ We know the problem well, but we still can’t agree on how to measure it.
The AI side has a clear example: the Epic Sepsis Model, used in hundreds of U.S. hospitals, was checked in 2021 and showed a sensitivity of 33% and a positive predictive value of 12%—so most alerts were false positives. That’s tough for a tool meant to catch a condition where speed is everything.⁴ It was meant to catch something life-threatening but ended up teaching clinicians not to trust the alarm.
Why the conversation matters more than the screen
I think we miss something when we treat this as just a design issue. Things like grouping alerts, ranking severity, or improving visuals are helpful, but they weren’t the real problem in my opening story. The issue started before any screen was designed. The team focused on reducing liability built exactly that: an alert that catches everything, no exceptions. No one in the room asked whether a clinician could handle it a hundred times a shift. No one had to be the person overriding it at 2 a.m.
That’s not a design flaw. Design didn’t have the power to fix it later—the blanket rule was already set before design began. It’s a governance issue: a choice about who gets to release an automated clinical tool, made without bringing in clinical judgment where the rules were decided.
It helps to point out how this differs from the usual AI-in-healthcare story—where clinicians trust a model too much and skip double-checking. That’s automation bias, and it’s a well-known issue. This situation is more like its flip side. Alert fatigue isn’t about trusting a signal too much; it’s about losing trust entirely because there’s so much noise that clinicians can’t tell which alerts matter. Automation bias asks: what if people believe the AI too easily? Alert fatigue asks: what happens after they’ve learned, rightly so, not to? Both come from launching an AI tool without a clear sense of how much information a person can handle—one from too little doubt, the other from too much. Better wording on the alert box won’t fix either.
The guardrail is a staffing choice
If there’s a design fix here, it’s not a new template for a library. It’s a rule about who gets to approve an automated alert before it goes live: not just legal or the team worried about liability, but someone who will actually deal with every false alarm it creates. That’s tougher to set up than a design review because it gives veto power to the person least likely to be in the room—especially when the fix is rushed after something already went wrong.
The lawsuit in this story wasn’t wrong to ask for a solution. The urge to close the gap wasn’t wrong either. What was missing was someone who could have said, before launch: this will be tuned out into silence, and we’ll just face the same issue again, only quieter.
Sources
Poly TN, Islam MM, Yang HC, Li YCJ. Appropriateness of overridden alerts in computerized physician order entry: systematic review. JMIR Med Inform. 2020;8(7):e15653. https://medinform.jmir.org/2020/7/e15653
Ancker JS, Edwards A, Nosal S, Hauser D, Mauer E, Kaushal R; HITEC Investigators. Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. BMC Med Inform Decis Mak. 2017;17(1):36. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5387195/
Ray CE, Wilson GM, Hughes AM, et al. Alert fatigue measurement in clinical decision support: a systematic review. J Am Med Inform Assoc. 2026;33(8):1523-1531. https://pubmed.ncbi.nlm.nih.gov/42148822/
Wong A, Otles E, Donnelly JP, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern Med. 2021;181(8):1065-1070. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307