Cancelled AI gun alert still sent armed police to a Maryland school
An AI-based weapon detection system flagged a student at Kenwood High School in Maryland. Two minutes later, the school's security team cancelled the alert. Seventeen minutes after the cancellation, armed police arrived anyway. The entire chain of events is now the subject of a lawsuit that could shape how schools across the country use similar systems.
The case filed
On November 17, 2026, the lawsuit Allen, et al. v. Omnilert, et al. was filed in the Circuit Court for Baltimore County in Maryland. The plaintiffs are four current and former students at Kenwood High School: Taki Allen, Yui-Osa Ogbeide, Devon Thompson Jr. and Keith Winfield III. They are represented by the Lawyers' Committee for Civil Rights Under Law, with Willkie Farr & Gallagher LLP as pro bono counsel and Engel Law Group, P.C. (Let's Data Science)
The plaintiffs accuse Omnilert, the school district Baltimore County Public Schools (BCPS), county police and individual officers of negligence and constitutional violations. They seek an injunction against further use of Omnilert in schools until protocols exist that prevent police responses to cancelled alerts, in addition to damages — the amount is not specified in the available sources (Ars Technica).
What actually happened, minute by minute
The incident took place on October 20, 2025, after football practice. According to the lawsuit, four young men were sitting together sharing a bag of Cool Ranch Doritos while waiting for a ride home, when the system flagged one of them as a potentially armed person (Ars Technica).
The timeline from BCPS, as reported by Newsweek, looks like this (Newsweek):
- 7:04 p.m. — Omnilert issues an alert from Kenwood High School.
- By 7:06 p.m. — designated security staff and a police representative have reviewed the images, concluded there was no threat, and cancelled the alert — a notification that, according to Ars Technica, informed both school leadership and Baltimore County police that no response was needed.
- 7:17 p.m. — principal Catherine "Kate" Smith contacts the school resource officer after seeing the original alert.
- 7:23 p.m. — the first officer arrives.
The lawsuit states that at least six police vehicles arrived with guns drawn, and that the young men were handcuffed and searched. These allegations are not verified; they come from the complaint, not from an official investigation (Ars Technica).
What actually triggered the alert?
The most widespread framing of the case — that the system "mistook a Doritos bag for a gun" — is, according to Maryland's Office of the Inspector General for Education, not accurate. In its November 2025 investigation, the inspector found that the system did not simply confuse a blue chip bag with a weapon. The alert was triggered by a combination of the student's hand movements, the object's position, and its reflective properties. The lawsuit states that the system calculated a 90 percent probability of a firearm (Newsweek).
That detail matters for how the failure is understood. A classification based on movement and reflection is closer to how visual detection systems actually work: they do not assess objects in isolation, but patterns in the image that statistically resemble weapon handling. It also makes the failure harder to fix with simple training data — a chip bag held in a particular way, in a particular light, can sit close to the images the model associates with firearms.
Conflicting numbers: 99 percent versus 96 percent
The lawsuit claims that over 99 percent of alerts in Baltimore County's schools were false positives, citing the inspector's investigation. But the inspector's own findings are more nuanced, and the figures measure different things (Newsweek):
- Omnilert's monitoring layer filtered out 96 percent of the objects the software initially flagged, without requiring further action.
- Of the alerts forwarded to BCPS in the period November 2023 to November 2025, fewer than 1 percent required police notification.
- And crucially: fewer than 1 percent does not mean the software misidentified a firearm in 99 percent of cases. Some of the forwarded alerts involved replica weapons or objects that warranted closer inspection.
The difference is that "false positive" in the lawsuit's framing covers everything that was not a real weapon in the hands of a threat, while the inspector's figures distinguish between the software's flags, the human monitoring layer's filtering, and which alerts ultimately triggered police contact. The lawsuit's choice of framing and the inspector's more categorized findings point in opposite directions — and that tension is one of the core conflicts of the case.
Where the human chain broke down
However the numbers are weighed, it was not the software alone that sent armed officers to the school. The inspector found weaknesses in the school's procedures that allowed already-cancelled alerts to be relayed to police as potential active threats, and recommended procedural changes and training. Both BCPS and the police acknowledged the need for improvement (Newsweek).
The exact human failure chain is disputed, however. The account associated with the Lawyers' Committee describes both technological and human errors, claiming that school staff and police attempted to cancel the alert, but that inadequate protocols meant armed police were dispatched anyway (Ars Technica). BCPS's timeline points to something else: that Principal Smith contacted the resource officer at 7:17 p.m. — eleven minutes after the cancellation — after seeing the original alert. Whether she failed to act on the cancellation notification, and why, are allegations in the complaint, not verified facts.
Statements attributed to the officers during the incident — including comments about "stupid AI cameras" and warnings about chip-bag objects — are likewise documented only through the plaintiffs' account and must be treated as such.
Omnilert's position — and the marketing question
Omnilert does not dispute the technical pattern of facts, but the interpretation. A company spokesperson told CBS Baltimore that the system "operated as designed, routing a possible threat for human review" (Let's Data Science). In other words: the alert was flagged and routed onward, humans reviewed it, and humans cancelled it — within two minutes. That something went wrong afterwards belongs, in this interpretation, to the school's and police's processes, not the product.
The plaintiffs' counter targets precisely the seam between product and process. They claim Omnilert markets a defective product with promises of "near-zero false positives" — a promise that in practice depends on human review that the process did not ensure (Ars Technica). It is a familiar challenge for the entire human-in-the-loop model: when a vendor communicates error rates from the automated layer, but actual safety depends on a human layer with its own error rates, who bears responsibility for the compounded harm?
BCPS has previously described its Omnilert rollout as the largest of its kind for visual AI weapon detection — a framing that gives the case reach beyond a single district (Ars Technica).
Damon Hewitt, president and executive director of the Lawyers' Committee, summarized the plaintiffs' diagnosis this way: "Overreliance on artificial intelligence technology, mixed with inept communication and coordination among key officials, is a toxic mix" (Ars Technica).
What is at stake
The case is the first major legal test — at least the first of its scale in the available sources — of what happens when a cancelled AI alert still leads to an armed police response. The precedent it could set concerns not what algorithms are allowed to flag, but how institutions must organize downstream of them:
- The injunction. The plaintiffs want BCPS to stop using Omnilert until protocols that actually prevent police responses to cancelled alerts are established. The court's stance on that will say a great deal about where responsibility is placed between vendor and institution.
- The protocol question itself. The inspector's findings point toward a concrete technical fix: cancelled alerts must be able to reach both school leadership and police in a way that cannot be bypassed — for example, with an automatic cancellation signal to the police response system, or a requirement of human verification before armed deployment against a flag without a new incident report.
- Liability. Whether the constitutional claims against the officers and the negligence claims against Omnilert survive motions to dismiss will be shaped by precisely the question of whether the cancellation failure was a protocol breakdown or an individual error.
Several questions remain open. No full text of the complaint or of the inspector's November 2025 report has been published in the available material; all the figures come from the plaintiffs' account or secondary coverage of it. Omnilert has commented only through the spokesperson's statement to CBS Baltimore. And earlier complaints against Omnilert — including cases tied to Antioch High School — are referenced in the coverage, but not documented in detail here.
The most concrete picture for now is this: the software largely performed its defined task — to flag, not to judge. A human system did the same, faster than expected. Yet it still ended with six police vehicles and handcuffs, according to the lawsuit. It is the gap between those layers, not the algorithm's error alone, that a court in Baltimore County must now decide.

