What happened

On December 9, 2025, Lawton Chiles Middle School in Oviedo, Florida went into a Code Red lockdown. A Code Red is the most serious lockdown level a school can call: doors locked, lights off, students herded into corners away from windows, the assumption being that there is an active threat inside the building. Police rolled in. For a stretch of that morning, an ordinary school day was being run as if a shooter was on campus.

There was no shooter. There was a student holding a clarinet.

The school district, Seminole County Public Schools, uses an AI weapon-detection system from a Pennsylvania company called ZeroEyes. The system watches the district's existing security camera feeds and is trained to spot the visual signature of a firearm, then push an alert to humans who decide what to do next. On that morning, a camera caught a student holding his clarinet up to his shoulder. As local reporting described it, he was holding the instrument "in the position of a shouldered rifle." He was also, as it happened, dressed in camouflage and tactical-style gear as part of a themed dress-up day at the school.

So the picture the AI saw was a person in military camo raising a long, dark object to their shoulder and pointing it. To a model trained to recognize the shape and stance of someone wielding a rifle, that is close enough. It fired the alert. The lockdown followed. Only after officers and staff reviewed the actual camera footage did anyone confirm that the "weapon" was a woodwind.

The part that should worry you most

The natural assumption, reading a headline about AI mistaking a clarinet for a gun, is that the technology malfunctioned. It did not, at least not by its vendor's definition. That is the genuinely unsettling part of this story.

Seminole County's superintendent said the safety system "worked as intended." A ZeroEyes co-founder, Sam Alaimo, went further and defended the flag directly, saying the image showed a student appearing to aim the clarinet like a gun at a door and that it strongly resembled, in his words, a shooter about to do something bad. From the company's point of view, the system was shown a person in camo shouldering an object and aiming it, and it raised a flag. That is exactly the behavior the district paid for.

This is the uncomfortable core of perimeter gun-detection AI: a model tuned to catch a real rifle from a low-resolution security camera, through bad lighting, at a distance, partially hidden by a body and clothing, has to treat anything rifle-shaped and rifle-posed as a candidate. A clarinet held at the shoulder by a kid in a tactical vest is, in pixels, not far off. You cannot dial the sensitivity down far enough to never flag a clarinet without also risking that you fail to flag an actual gun. The false positive is not a bug in the model. It is the model doing its job in a world where guns and clarinets can briefly look alike on a hallway camera.

Which means the interesting failure is not "the AI got it wrong." It is "what the district had wired up to happen the instant the AI got it right by its own standard."

Why this still belongs in a graveyard of AI mistakes

If a vendor exec can stand up and defend the alert as correct behavior, is this even an AI failure? Yes, and here is the distinction worth being precise about.

The model's per-frame judgment may have been defensible. The system, as deployed, was not. Seminole County paid roughly $250,000 for ZeroEyes and connected its output to a chain that escalated a single camera flag into a full Code Red and a police response before a human had confirmed what the object actually was. The consequence of a false positive scales entirely with what you bolt to the other end of the alert. Connect it to a security officer who quietly checks the feed, and a clarinet costs you thirty seconds. Connect it to an automatic lockdown and a 911-style police mobilization, and a clarinet costs you an entire school's morning, a building full of frightened kids, and armed officers responding to a threat that was never there.

That is the teachable failure here, and it is an AI-deployment failure even if the inference was "correct." When you put a probabilistic detector at the front of a high-stakes, low-tolerance response protocol, every false positive the model is guaranteed to produce becomes a real-world emergency. The vendor gets to say the tech worked. The students get to experience the lockdown. Both statements are true at once, and that gap is precisely the problem.

The other side of the ledger deserves an honest hearing too. Defenders of these systems argue, not unreasonably, that you would rather have a thousand clarinet lockdowns than miss the one real rifle. School shootings are a genuine and recurring horror, and the appeal of a camera that never blinks is obvious. The catch is that a detector which cries wolf often enough trains everyone, students, staff, and police, to treat the next alert as probably-another-clarinet. The ACLU's Chad Marlow made roughly this point in coverage of the incident: systems that throw frequent false alarms can manufacture a false sense of security while subjecting children to repeated, traumatic lockdowns and policing. A boy-who-cried-wolf detector does more than annoy; it erodes the response it exists to trigger.

A documented pattern, not a one-off

This was not an isolated freak event. AI weapon-detection in American schools has a growing track record of confidently flagging the wrong thing, and occasionally missing the right thing.

The most direct comparison is a separate incident from October 2025, in Baltimore County, Maryland, where a different vendor's system (Omnilert) flagged a high school student's bag of Doritos as a firearm. In that case a human reviewer had actually canceled the alert, and a school administrator called police anyway; the student, a Black teenager holding a snack, was handcuffed and searched at gunpoint. [1] The two incidents are genuinely distinct: different vendor, different state, different object, different downstream failure. But they rhyme. Coverage of the clarinet lockdown even noted that ZeroEyes had previously been involved in mistaking a bag of chips for a firearm, underscoring that "rifle-shaped object held by a person" is a category the models keep over-including.

Stack these together and a clear lesson emerges. The technology will keep producing false positives, because that is mathematically what a sensitive detector does. The question that determines whether anyone gets hurt is the human-and-policy layer: who reviews the flag, how fast, with what authority to stand it down, and what is allowed to happen automatically before that review completes. In Baltimore the review worked and a human overrode it anyway. In Oviedo the lockdown apparently preceded the confirming human review entirely. Same root vulnerability, two different ways to get burned.

The transparency footnote

There is a smaller, quieter failure in the Seminole County story that is easy to skip past. According to the local reporting, the district's public statements after the incident emphasized that the system "worked as intended" but omitted the actual context: the clarinet, the costume, the dress-up day. A parent quoted in the coverage pointed out the obvious consequence: with the details left out, the community was left believing the AI had simply malfunctioned and flagged a random kid, when in fact the picture was a camo-clad student shouldering and aiming an object. The omission did the technology no favors. It turned a defensible-if-overcautious detection into what looked like a clown-show glitch, precisely because nobody explained what the camera saw.

That is its own lesson about deploying surveillance AI in public institutions. The trust you need for these systems to be tolerated at all depends on telling people what actually happened when they go wrong, including the embarrassing specifics. "It worked as intended" is not reassuring when the intended behavior is locking down a middle school over a band instrument.

What it teaches

The clarinet incident is, mercifully, a near-miss rather than a tragedy. No child was held at gunpoint this time. But it is a clean illustration of how AI safety systems fail in practice, which is rarely a dramatic crash and usually a confident, "working as intended" output plugged into a response that does not match the output's reliability.

The hardware did what it was sold to do. The model did what it was trained to do. And a school full of kids still spent a December morning in a Code Red because a 12-year-old in a costume picked up his instrument. If you are going to put a probabilistic detector in front of armed-response infrastructure, the detector's accuracy is almost beside the point. What matters is whether a human, with the authority and the time to say "that's a clarinet," sits between the alert and the lockdown. In Oviedo, that human apparently came in second.

[1] If you want the gory details of the Doritos case, it has its own entry in this graveyard. Different vendor, same genre of failure.