On 17 November, four current and former students at Kenwood High School in Baltimore County filed a lawsuit in the county’s Circuit Court. The defendants are Omnilert, the company that sells the AI gun detector the school district runs on its cameras, along with Baltimore County Public Schools, the school’s principal, the district’s head of school safety, the county police department, five named officers and three unnamed ones. The complaint describes the evening of 20 October 2025. Taki Allen and three teammates were waiting outside after football practice, and Taki was eating a bag of Cool Ranch Doritos his coach had handed out. The camera system flagged him as carrying a gun. At least six police cruisers arrived, officers drew their weapons, and all four students were made to kneel on the concrete, handcuffed and searched. No weapon was found.
These are allegations, and the defendants haven’t answered them in court yet. But a lot of the factual record was already public. Maryland’s Office of the Inspector General for Education investigated the incident and published a seven-page synopsis on 25 November 2025, and the complaint draws on it. Reading the two side by side, I think the complaint’s headline number tells us less than the timeline in its paragraphs 77 to 90.
The number in the complaint
The complaint says the system is unreliable and puts a figure on it: “Maryland’s Office of the Inspector General for Education (‘OIGE’) found that more than 99% of OGD Alerts in Baltimore County Public Schools were false positives that required no police response.” It adds that the system flagged the chip bag “with 90% confidence.”
The inspector general’s report doesn’t use the words “false positive.” What it says, in its second finding, is that over November 2023 to November 2025 the Omnilert monitoring system “filtered out 96% of the objects it alerted on and did not require further action,” and that “fewer than 1% of alerts forwarded to BCPS by Omnilert required notifying law enforcement.” It also describes the funnel those alerts pass through. The AI generates “thousands of alerts daily.” An initial review screens out things like Junior ROTC replica rifles and look-alike toys. Roughly 200 go to Omnilert’s monitoring staff for visual inspection, and after that the district receives “approximately 12 alert notifications daily” for its own safety staff and selected police to review.
The report’s own numbers don’t line up perfectly. Twelve out of 200 means 94 percent filtered at the vendor’s review, not 96, and it isn’t clear which stage the 96 percent covers. That’s minor. The bigger issue is what “fewer than 1%” measures. It counts alerts that “required notifying law enforcement.” That’s a strict bar. An alert on a JROTC drill rifle, or on a child with a toy gun, is a correct detection of something gun-shaped that doesn’t need police, and a system built to look for guns is going to produce a lot of those. Calling all of them false positives, as the complaint does, makes the detector sound worse at seeing than this measure can show. It does show something else: at 12 a day, fewer than 1 percent means the district’s reviewers go about eight days of alerts, and probably much longer, between alerts that need a police response.
That ratio is normal for alarms
A stream that’s more than 99 percent non-actionable sounds damning. In the literature on alarms, it’s ordinary.
The closest comparison I know is hospital monitoring. William Fleischman and colleagues had an emergency physician spend 53 hours watching the monitors in an urban academic emergency department and record every alarm, published in the American Journal of Emergency Medicine in 2020. There were 1,049 alarms on 146 patients, a median of 18 an hour. “Alarms changed clinical management in 8 out of 1049 observed alarms (0.8%, 95% CI, 0.3%, 1.3%).” Staff didn’t visibly respond to 63 percent of them. That’s almost exactly the Baltimore County ratio, from a field that has built its whole alarm-safety program around the problem.
Air traffic control is the counterexample in the other direction. Christopher Wickens and colleagues, in Human Factors in 2009, analysed 495 mid-air conflict alerts from five en route centres and found that 45 percent were false, ranging from 28 to 58 percent across centres. So more than half were real. A gun detector watching school cameras can never get near that, because guns on school grounds are so rare. It’s the same arithmetic I went through for AI-text detectors in /writing/zero-false-positives-in-twenty-five/: when what you’re looking for is rare, even a small error rate per look fills the alert queue with innocent cases.
That’s also why the “90% confidence” in the complaint doesn’t mean much. A detector’s confidence score is a number the model outputs, and the operator picks a threshold on it. Benchmark scoring hides how that choice plays out. Simon Wenkel and coauthors, writing in Sensors in 2021, point out that to score well on benchmarks, most neural networks use “a rather low threshold as a high number of false positives is not penalized by standard evaluation metrics.” Whatever the score on the chip bag was, the share of Baltimore County alerts that needed police was under 1 percent. A school district should plan around that number, not the score.
The humans were built to say yes, except one group
Omnilert says outright which way it leans. The complaint quotes the company saying that “visual gun detection systems are intentionally designed with a clear priority: not missing a real firearm,” and that false positives are “a result of system design choices and workflow ownership, not just detection accuracy.” I agree with the second statement more than the company might like.
The first human layer is Omnilert’s own reviewers, and their instructions push them the same direction as the model. Left alone, people searching for rare targets lean the other way. Jeremy Wolfe, Todd Horowitz and Naomi Kenner showed in Nature in 2005 that in an artificial baggage-screening task, observers missed 7 percent of targets at 50 percent prevalence and 30 percent at 1 percent prevalence, while false alarms were “vanishingly rare (0.03%).” Left to themselves, people looking for rare things mostly say no. Omnilert’s design turns that around on purpose, and on 20 October its reviewers passed on a picture of a teenager holding a chip bag.
The district’s reviewers were the second human layer, and they’re the part of the story that worked. According to the complaint, the alert went out shortly before 7:04 pm to at least fifteen people, five at the district and ten at the county police. Between 7:05 and 7:06, the district’s executive director of school safety, its school safety manager and a police sergeant looked at the image, saw a chip bag and canceled the alert. A cancellation text went to everyone who had received the original.
That fits the more hopeful part of the alarm literature. In James Bliss and colleagues’ 1995 experiment in Ergonomics, “most subjects (about 90%) do not respond to all alarms but match their response rates to the expected probability of true alarms,” which is the classic cry-wolf result. The air traffic controllers in the Wickens study didn’t show it, though. They often changed an aircraft’s trajectory before the alert even fired, which the authors read as controllers watching the raw data alongside the system rather than waiting for it. The district’s reviewers had the same advantage. The alert came with the image, and they used it in under two minutes.
Fifteen minutes after the cancellation
Then the alert came back.
The complaint says that between 7:17 and 7:19, the principal opened the original text, followed its link to a page that “expressly indicated the Alert had been canceled,” enlarged the video looking for a weapon, took a screenshot and sent it to the school resource officer on his personal phone. At 7:19 she texted two district administrators: “Omni alert is working. Kid standing on my patio right now. Looks like one of mine and looks like he’s armed. My SRO is calling it into the precinct now.” The officer forwarded the screenshot to the precinct. Officers were dispatched at 7:21 to a “report of a suspicious person with a weapon,” about fifteen minutes after the cancellation went out.
The principal’s account to the inspector general is less damning. She said she paused and enlarged the video, tried to call the zone safety manager and got no answer, and noticed the cancellation only after police had cleared the grounds. The report’s sixth finding supports her on the interface: “zooming in on the red-boxed still image or watching the embedded video could cause the user to overlook the cancellation text message when the photos are enlarged.”
Either way, the structural finding is the same. The district had a rule about who could cancel an alert, and only the executive director of school safety had “the primary authority to issue a cancellation override of a cancelled alert.” But the inspector general found that “other personnel receiving an alert could bypass or circumvent it by simply making a phone call or sending a text message to BCoPD personnel.” Its eighth finding notes the same pattern at Parkville High School on 7 November 2025: the district canceled, and outside recipients kept the response going out of “an abundance of caution.” The report doesn’t say whether police were dispatched in that case.
That’s what fifteen recipients does to a stream that’s more than 99 percent noise. Each extra recipient is another person who can turn a canceled alert back into a live one with a phone call, and nobody stops a teacher, principal or police officer who wants to be careful about a possible gun. The 1995 cry-wolf finding warns that people will learn to ignore alarms. In this case one recipient saw an old alert as new, and nothing in the process was designed to stop that.
What the numbers say should change
The inspector general’s recommendations are about the workflow, not the detector. They are to cut the number of people who get the first alert, bring in police only through a second notification once the district confirms a threat, switch designated staff from text messages to Omnilert’s app so everyone sees the same view, and train staff and police twice a year. Each of those reduces the number of ways a canceled alert can come back.
I’d add one thing, and it comes from the funnel. The district pays $560,000 a year for the monitoring service, which the report works out to $5.09 per student across 110,066 students. For that, it gets about 12 alerts a day for human review, and police are needed on fewer than 1 percent of them. Any process built on that stream should assume every alert is probably wrong. A canceled alert should stay canceled everywhere it was sent, and a police dispatch should have to go back to the one person with authority over it.
References
- Allen v. Omnilert, LLC, et al. Complaint, Circuit Court for Baltimore County, Maryland, e-filed 17 November 2026.
- Maryland Office of the Inspector General for Education (2025). Investigative Synopsis, OIGE Case 25-0014-I. Issued 25 November 2025, amended.
- Fleischman, W., Ciliberto, B., Rozanski, N., Parwani, V. and Bernstein, S. L. (2020). Emergency department monitor alarms rarely change clinical management: An observational study. American Journal of Emergency Medicine, 38(6), 1072-1076, June 2020.
- Wickens, C. D., Rice, S., Keller, D., Hutchins, S., Hughes, J. and Clayton, K. (2009). False Alerts in Air Traffic Control Conflict Alerting System: Is There a “Cry Wolf” Effect?. Human Factors, 51(4), 446-462, August 2009.
- Wenkel, S., Alhazmi, K., Liiv, T., Alrshoud, S. and Simon, M. (2021). Confidence Score: The Forgotten Dimension of Object Detection Performance Evaluation. Sensors, 21(13), 4350, June 2021.
- Wolfe, J. M., Horowitz, T. S. and Kenner, N. M. (2005). Rare items often missed in visual searches. Nature, 435, 439-440, May 2005.
- Bliss, J. P., Gilson, R. D. and Deaton, J. E. (1995). Human probability matching behaviour in response to alarms of varying reliability. Ergonomics, 38(11), 2300-2312, November 1995.