AI Security Guards: Missing Guns, Flagging Doritos

Quackery

When I was a kid in the late 1970s and early 1980s, the chemical compound laetrile was a huge deal. While researching this piece, I did a search on Newspapers.com (worth every penny) for "laetrile" between 1975 and 1979, and got 97,901 matches.

According to the National Cancer Institute, "laetrile" is the common name for a derivative of the chemical compound amygdalin, a bitter-tasting substance found most commonly in the pits of stone fruits like peaches and apricots, as well as in the seeds of cherries and apples, and in almonds. The U.S. Food and Drug Administration describes amygdalin as a "natural toxin":

The chemical is not found in the fruit itself and accidentally eating a seed or pit will not harm you. However, consuming a large amount of the seeds or pits can be problematic because enzymes in your intestines can turn amygdalin into cyanide and cause cyanide poisoning.

Notwithstanding its obvious risks, amygdalin ("laetrile," or even more deceptively, B-17) was a widely-touted treatment for cancer in certain circles. The John Birch Society, a noted far-right group, launched a campaign to promote pro-laetrile legislation in state legislatures in the mid-1970s. Much of the lobbying centered on individual rights and anti-government sentiment, but the movement was infested by neo-Nazism and anti-Semitic tropes as well. In many ways, laetrile was the ivermectin of its day.

Citing its lack of proven efficacy as an anti-cancer drug and its disturbing potential for cyanide poisoning, the U.S. Food and Drug Administration has long refused to approve it as a valid medical treatment or for use as a vitamin (it is colloquially known as B-17). Interest in laetrile has fallen dramatically; the compound has only been mentioned in a few dozen articles over the last six years.

My wife and I are the parents of a childhood cancer survivor. I have a visceral understanding of the fear that blows open the Overton window of medical treatment; you're willing to try anything that might help. We were relatively fortunate; our child had Hodgkin's lymphoma, a cancer that is well-understood and eminently treatable. The five-year survival rate is above 95%, and we are blessed to be in that majority.

I was also elected to the Burlington (VT) school board three years after the 1999 Columbine High School shootings. Burlington is nearly 2,000 miles from Littleton, Colorado, but the scale of the tragedy was not something that the media (or any parent of school-age children) could ignore. Again, we were fortunate not to have any shootings in our school system, but students would occasionally bring knives and other potentially harmful items to school. We regularly had conversations about how to balance student privacy rights with school safety.

The Fear Industry

More than two decades later, the school safety industry is monetizing parental fear in powerful new ways. It's a far too easy sell; since the turn of the century, there have been thousands of shooting incidents at schools around the country. Contemporary parents can recite the worst tragedies – Sandy Hook, Parkland, Uvalde – in much the same way that our grandparents listed Pearl Harbor, Iwo Jima, and the Battle of the Bulge.

Nobody wants to be the next name in that list. According to National Public Radio, schools already spend as much as $4 billion per year, and that figure is expected to grow steadily. A significant percentage of that growth will be driven by AI-powered products that purportedly can detect criminal behavior, electronically search backpacks for guns, and monitor students' social media posts and school browser activity in real time (more on the digital privacy issues in a later post).

AI security "guards," however, are just not ready for prime time. There are myriad examples of overly-aggressive algorithms misidentifying harmless items as guns. Dr. David Riedman, the founder of the K-12 School Shooting Database, maintains a growing list of examples of AI surveillance failures.

In December 2025, for instance, an AI system installed in the Sayreville (NJ) school district mistook a vodka bottle for a gun. The district pays $128,000 per year for the software.

That same month, the AI security system ZeroEyes (seems a bit on the nose, honestly) interpreted a middle school student's clarinet as a gun. The building was locked down, and police were called to deal with someone described as "a man in the building, dressed in camouflage with a 'suspected weapon pointed down the hallway, being held in the position of a shouldered rifle.'” A spokesperson for ZeroEyes effectively blamed the student, saying that he fooled the system by “intentionally holding the instrument in the position of a shouldered rifle.”

Far more disturbing was the false positive that occurred in Baltimore in October 2025. Following football practice, 16-year-old Taki Allen finished off a bag of Doritos and stuck the bag in his pocket while he waited outside for a ride. Omnilert, an AI gun detection company installed by the Baltimore County Public Schools, scanned Allen as he stood near the school and sent out a shooter alert. A short time later, Allen said, "Police showed up, like eight cop cars, and then they all came out with guns pointed at me talking about getting on the ground." The police handcuffed the teen, but reassured the public that they hadn't actually arrested him. But it was obviously a situation that could have gone sideways far too easily. Allen told local media that he no longer waits outside.

False positives are bad enough, but false negatives can be tragic. On January 22, 2025, a 17-year-old student at Antioch High School in Antioch, Tennessee (near Nashville), shot and killed a female student and wounded two others. He then fatally shot himself.

Six months later, the parents of the slain student filed a lawsuit against the Metro Nashville Public Schools (MNPS) and Metro Government, seeking $700,000 in damages (the maximum damages under Tennessee law). The lawsuit alleged that the defendants were aware of the shooter's propensity for violence and failed to take action. It also alleged that the Omnilert system installed by MNPS failed to detect the shooter's weapon (although the system did issue an alert when police responded to the shooting with drawn guns). That lawsuit was settled in November 2025 for $300,000.

Last month, one of the injured students sued Omnilert and a Tennessee security company, alleging that the AI system had design defects, that Omnilert overpromised what the system could do, and misrepresented its capabilities. The plaintiff points out that shortly after the Antioch shooting, Omnilert changed the language on its website to soften its marketing promises. A copy of the complaint can be downloaded below.

Misdirected Resources

Notwithstanding the clear failure of the AI algorithm to detect a weapon threat, MNPS voted in February to pay $1.25 million per year to Evolv Technologies, a Waltham, Massachusetts company that offers "Advanced Security Detection Technology for a Safer World."

Riedman, who worked briefly for ZeroEyes, questions MNPS's decision. That money, he told ArsTechnica, “could have gone to a counselor or something else to a kid in crisis. Every decision that you make is pointing away resources from something else.”

It will take a courageous school board, in the current environment, to choose expanded social services over automated systems, even ones with a spotty track record. But they should do so. The alternatives are grim: Either these systems don't do what is promised (in which case kids are put at risk), or they do it so well that they obliterate staff and student privacy. I worry about the former, but the latter scenario is truly Orwellian.