The most damning fact in the new Baltimore County lawsuit against Omnilert is not that the gun-detection AI was wrong. It is that the AI was wrong, everyone knew it was wrong, the alert was formally cancelled, and armed police still put four Black teenagers on their knees in handcuffs outside Kenwood High School. The Lawyers’ Committee for Civil Rights Under Law filed the suit on November 17 in the Circuit Court for Baltimore County, naming Omnilert, Baltimore County Public Schools (BCPS) and personnel, the Baltimore County Police Department, and individual officers. The plaintiffs are Taki Allen, Yui-Osa Ogbeide, Devon Thompson Jr., and Keith Winfield III.
Read the timeline the complaint lays out and the story stops being about computer vision. Omnilert Gun Detect, deployed across Baltimore County schools, flagged a student’s bag of Doritos as a firearm with a stated 90% probability in October 2025. School staff and local police reviewed the image and determined it was a mistake. The alert was cancelled. Armed officers were dispatched anyway, because, per the Lawyers’ Committee, school and police officials “failed to have sufficient protocols in place to prevent cancelled alerts from triggering armed police responses.” That is the mechanism. Not a model hallucinating in a vacuum. A human override that did not actually override anything.
What the lawsuit actually alleges
The complaint alleges violations of the students’ federal and state constitutional rights, negligence, and gross negligence. The requested relief is procedural rather than punitive in framing: the Lawyers’ Committee wants local police and school procedures reformed so public actors verify AI threat reports, evaluate whether this class of technology belongs in schools at all, and build policies ensuring a false alert cannot summon armed police. It also asks for training, oversight, transparency around errors, and community input on deployment.
The specific defendants and the specific ask matter more than the general outrage. This is not a suit against “AI” as an abstraction. It is a suit against a named vendor, a named school district, and a named police department, over a named product with a documented tendency to misfire. The Lawyers’ Committee states that school officials “knew” the system “regularly produced false alerts.” If that holds up in discovery, Omnilert’s exposure is not a one-off bug. It is a known-defect claim.
Damon Hewitt, president and executive director of the Lawyers’ Committee, framed it as a systems failure: “Overreliance on artificial intelligence technology mixed with inept communication and coordination among key officials is a toxic mix.” Leah Frazier, who directs the organization’s Digital Justice Initiative, put the escalation question bluntly: “This should have never been escalated to the police.”
The 90% number is the tell
A 90% confidence score sounds authoritative to a principal or a dispatcher. It is not a probability that a gun is present. It is, at best, a model’s internal score on a cropped image, calibrated against training data that has no idea what a foil chip bag looks like crumpled in a backpack at an oblique angle under fluorescent light. Vendors rarely publish calibration curves, false-positive rates by demographic or setting, or the operating thresholds they recommend. Omnilert Gun Detect is sold into schools on the promise of catching shooters in seconds. The failure mode, as this case shows, is a pipeline where a high score from an uncalibrated detector becomes a dispatch decision with almost no friction.
The racial dimension is not incidental. The Lawyers’ Committee notes that harms from these systems “disproportionately impact Black communities and other people of color,” and the four plaintiffs are Black students at a majority-Black school. The burden of a false positive is not distributed evenly across a district. It lands on whichever kids happen to be standing near the flagged object when the alert fires.
The AI was wrong, everyone knew it was wrong, the alert was cancelled, and armed police still cuffed four Black teenagers.
What this means for the people building this stuff
Three things, and none of them are comfortable for the computer-vision vendors selling into public safety.
First, a confidence score is not a decision. Any system that outputs a threat probability into an operational channel needs a downstream gate that a human can actually close. The Baltimore County failure was a cancelled alert that still dispatched armed officers. That is an integration and policy bug as much as a model bug, and it is the kind a vendor can fix with a single well-documented webhook contract: cancellation must propagate to dispatch, or the deployment should not exist.
Second, “we told them it was a demo” or “the customer misconfigured it” will not survive a negligence claim when the vendor’s own staff are alleged to have missed the same evidence the school did. The complaint describes a chain of human failures spanning Omnilert staff, the principal, and BCPD. A vendor that markets to schools inherits a duty of care proportional to the harm its false positives can trigger. Handcuffs and drawn weapons are the foreseeable output.
Third, the regulatory vacuum is the business model. Frazier’s line, “Unregulated AI has profound consequences,” is the whole thesis. There is no federal standard for weapon-detection accuracy in schools, no required disclosure of false-positive rates, no audit trail requirement, no calibration reporting. Vendors compete on sensitivity (catch every gun) and quietly absorb the specificity cost (flag every chip bag) until a lawsuit makes it public.
The case is Allen, et al. v. Omnilert, et al., filed in the Circuit Court for Baltimore County. Discovery will pull the model’s evaluation data, the deployment contract, and the dispatch logs. What those documents show about how many prior false alerts BCPS logged, and what Omnilert told the district about them, is the next thing to watch.