Four Students Sue After an AI Gun Alert Mistook a Doritos Bag for a Firearm
Posted on 20th Nov 2026 11:03:26 in Artificial Intelligence, Machine Learning
Tagged as: AI governance, AI surveillance, Omnilert, automated decision making, AI safety, public sector AI
On the evening of 20 October 2025, a group of football players at Kenwood High School in Baltimore County, Maryland, finished practice, took bags of Doritos from their coach and sat near the picnic tables waiting for their rides home. Within minutes, six police cruisers arrived with lights flashing. Officers stepped out with weapons drawn. The teenagers were ordered to their knees on the concrete, handcuffed and searched.
The tip-off was not a witness, a threat or a phone call. It was an artificial intelligence system called Omnilert Gun Detect, which scanned school security footage and flagged a bag of Cool Ranch Doritos as a probable firearm with 90 percent confidence. More than a year later, on 17 November 2026, four current and former students filed suit in Baltimore County Circuit Court against the company that built the system, the school district, the police department and individual officials. The case, reported this week by The Guardian, USA Today and Ars Technica, is already being treated as a test case for one of the hardest questions in enterprise AI: when an automated system raises an alarm, who is accountable for what happens next?
What Happened Outside Kenwood High School
According to the complaint, the students were sharing chips after practice when the Omnilert system — which continuously analyses camera feeds for the visual signature of firearms — generated an alert. School staff and a police representative reviewed the image almost immediately and concluded the system had made a mistake. The alert was marked resolved and cancelled, with notifications sent to everyone on the distribution list.
That should have been the end of it. It was not. The suit alleges that the school's principal, who had never been trained on the Omnilert system according to a state inspector general's report, noticed the original alert without registering the cancellation. Roughly fifteen minutes later she texted a school resource officer asking for police to be deployed — while acknowledging in the same message that the object in the image was "maybe not" a gun.
What followed, per the complaint, was a full armed response: at least six patrol cars, officers with guns drawn, students separated, forced to their knees, handcuffed and searched. The plaintiffs — Taki Allen, Yui-Osa Ogbeide, Devon Thompson Jr. and Keith Winfield III — say officers told them nothing about what they were suspected of until after no weapon was found. One officer allegedly muttered "stupid AI cameras" and "fucking Omnilerts" as he left. Another told the group, "Hey, thanks for not running away, that would have been a real problem." A colleague joked that the boys should be careful "with your chip bags."
The lawsuit describes handcuff marks as "deep red marks" on wrists, along with "acute and lasting emotional distress, fear, anxiety" and humiliation. Taki Allen, the student flagged as the gunman, no longer waits for rides outside school and now frequently pays for private car services instead. Keith Winfield's home is close to campus, but his mother Frieda says his anxiety was severe enough that he called her repeatedly after practice to confirm she was on the way.
The Cancellation That Never Reached the Cruisers
Strip away the novelty of the Doritos bag and the incident follows a familiar pattern in automation failures: the machine raised a false alarm, a human being correctly overrode it, and then the override was lost in the handoff to the next system.
Omnilert's own account supports this reading. The company told USA Today that the school's designated safety staff reviewed the alert and cancelled it within about two minutes, that everyone who received the alert was notified, and that the case demonstrates "why human review matters, and why a canceled alert must stay canceled." In other words, the vendor's position is that the product worked and the human process around it did not.
The Maryland Office of the Inspector General for Education reached a similar conclusion in its review. It found that school staff and a police representative determined the alert was not credible and sent a cancellation notice — and that the incident escalated because a school employee missed the cancellation and forwarded a screenshot to a law enforcement officer. The office could not substantiate claims that the Omnilert system was ineffective, inefficient or biased, and attributed the outcome to "human error and training gaps." It also noted that the software did not literally classify a chip bag as a gun: the alert was triggered by the student's posture, hand movements and the reflective properties of an object in the frame.
That last detail matters more than it first appears. The system was not wrong in the way a headline suggests. It was wrong in the way most production AI is wrong — a pattern that resembled a threat closely enough to cross a threshold, in a system tuned deliberately to over-flag, with the burden of disproof placed on people who had not been trained to carry it.
What the Lawsuit Alleges and What It Asks For
The complaint names Omnilert, Baltimore County Public Schools, the principal, the Baltimore County Police Department and individual responding officers. The plaintiffs allege unreasonable search and seizure, false arrest, false imprisonment and negligence, and argue their constitutional rights were violated when they were detained and searched on the basis of an alert that had already been withdrawn.
Beyond damages, the students are seeking an injunction barring the district from continuing to use Omnilert until it establishes protocols and training sufficient to prevent police responses to false alerts. Their legal team — the Lawyers' Committee for Civil Rights Under Law, Willkie Farr & Gallagher and Engel Law Group — says it wants verification requirements before any AI threat report escalates to an armed response, plus clearer oversight and transparency around how these systems are configured.
Damon Hewitt, president and executive director of the Lawyers' Committee, framed the case as a warning rather than a one-off: "These students were failed not only by a technology that is prone to false alerts, but also by the very adults meant to keep them safe when we send them to school." He described overreliance on AI mixed with poor communication between officials as "a toxic mix."
The Numbers Behind "Near-Zero False Positives"
The complaint's sharpest challenge is to the marketing. It alleges Omnilert sold the district on "near-zero false positives" without effectively requiring the human review that claim depends on, and notes that Baltimore County Public Schools was among the earliest adopters, with officials describing it as "the largest deployment of AI visual gun detection."
The lawsuit also cites Maryland's education inspector general as finding that more than 99 percent of Omnilert alerts in Baltimore County schools were false positives that required no police response. Omnilert disputes the framing, pointing out that the inspector general found no bias in the system and that the vendor's stated design goal is to surface anything that could be a firearm for human review, on the argument that "missing a real gun is the worst possible outcome."
Both positions can be true at once, and that is precisely the problem for buyers. A system tuned to maximise recall will fire constantly; an operator who sees 99 useless alerts for every real one learns, quite rationally, to discount them. The failure is not that the model is inaccurate. It is that an accuracy figure was treated as an operational guarantee, with no design work done on what happens to a signal between the alert and the armed response.
Why This Matters for Every Business Running AI in the Loop
It is tempting to file this under American school-safety politics. That would be a mistake. The same architecture — camera or sensor, model, alert, human reviewer, downstream action — is now standard in Indian retail loss prevention, warehouse safety monitoring, insurance claims triage, banking fraud flags and factory quality control. The failure mode travels with the architecture, not the industry.
Three lessons stand out for teams deploying monitoring AI in India, where facial recognition and video analytics are spreading through campuses, housing societies, shops and plants:
- Cancellation is a first-class workflow, not a status flag. The moment a reviewer marks an alert invalid, that decision must propagate to every downstream consumer with the same authority as the original alert. If a WhatsApp screenshot can outrun the official cancellation, your process is broken regardless of model quality.
- Only trained people should hold escalation rights. The report found the principal had no training on the system. Seniority is not competence when it comes to interpreting a probability score, and an untrained decision-maker with escalation authority is the single highest-risk node in the chain.
- Sell and buy recall and precision separately. "Fewer false positives" is not a specification. Ask for the false-alarm rate measured in the buyer's own environment, the review SLA, the audit trail for each alert's lifecycle, and a documented threshold change process. Then put those numbers in the contract.
There is also a governance point that Indian organisations tend to skip until an incident forces it: define, in writing, which automated alerts are allowed to trigger a physical or financial action without a second human confirmation. A gun alert dispatching armed officers is an extreme case, but the same question applies to an AI flag freezing a customer's account, blocking a delivery or triggering a disciplinary process.
What Good Alert Governance Looks Like
The practical test is simple. Take any AI alert your organisation acts on and trace it end to end: who receives it, what they are trained to do, how a rejection is recorded, who is authorised to escalate, and how the escalation is verified as legitimate before it takes effect. If any of those five answers is "it depends" or "whoever sees it first", the deployment is running on luck.
The Kenwood case will take years to resolve, and the legal findings may well split responsibility between vendor, district and police. But the operational finding is already available: the model produced a false positive, and a human corrected it. The harm happened in the gap after the correction — in a system designed to raise alarms but not to withdraw them quickly enough.
That gap exists in almost every AI monitoring deployment. It is inexpensive to close, and it does not depend on a better model. It depends on treating a human override as a hard system state, not a message someone might read.
Sources
- The Guardian — Four Maryland high school students sue after AI mistakes Doritos bag for gun
- USA Today — Baltimore students' lawsuit says AI system mistook Doritos bag for gun
- Ars Technica — Traumatized teens sue school, cops after Doritos bag triggers AI gun alert
- CBS News Baltimore — Baltimore County students sue over AI system that mistook Doritos bag for gun
- ACLU — Gun-Toting Police Swarm, Handcuff Young Black Man After AI Mistakes Doritos Bag For a Gun
- Lawyers' Committee for Civil Rights Under Law — Lawsuit announcement