Students vote to end AI surveillance contract

89.6% of student voters at William & Mary University decided to end the contract with an AI gun detection company that analyzes, stores, and perpetually owns data from campus security cameras.

Unlike k-12 schools, many colleges allow students to vote on decisions that impact their life on campus. This week, 1,256 students voted in a special ballot session to end the College of William & Mary’s $82,600 per year contract with the AI surveillance company ‘ZeroEyes’. The company claims they monitor CCTV cameras using AI to detect visible firearms in real time. It’s unclear if the company’s name is inspired by the eye on the cover of George Orwell’s 1984.

Per information in the voting referendum explanation, students cited three major concerns:

1. “ZeroEyes has not released data about the accuracy of their system”

2. “ZeroEyes cannot detect concealed weapons”

3. “The current contract grants this private vendor the perpetual and irrevocable license to use collected imagery for its own business operations”

William & Mary The Flat Hat student newspaper

In addition to the students’ concern, an administrator told the student newspaper the system triggered five false alarms in just 3 months of operation:

“I can tell you right now that we’ve had five notifications since we went live in January of this year,” Vice President for Public Safety Cliff Everton said. “All of them were what we would call a non-lethal notification, where somebody is walking on campus with a mock-up of a musket or a Nerf gun.”

The College of William & Mary in Virginia is one of the top 25 public universities in the country, so these are some very smart students who are highly attuned to thinking about social and political issues. As our entire society grapples with AI ethics, these students are rightly concerned about being under surveillance by a system with unproven performance and having a private company own videos/images of them living, learning, and socializing around their campus.

While the university was paying for this monitoring service, the video data the company gets to collect is a huge business of it’s own. AI companies are paying for videos of people doing real-world tasks to use for model training. When the contract grants this private vendor “the perpetual and irrevocable license to use collected imagery for its own business operations”, what stops a surveillance company from selling thousands of hours of CCTV footage from the campus for AI training? When AI is threatening entry level jobs, students should have a right to decide if videos of them on campus can be used and sold for AI training.

It also remains unclear whether AI companies are actually using AI or if they are using low-cost human reviewers/operators in place like Southeast Asia. For example, Amazon Go stories that were advertised as AI checkouts were actually humans watching cameras. The Waymo robo taxis also have a team of foreign workers who can take control if the cars get stuck but it’s unclear if/when remote humans are at the controls. Just like Amazon, humans could actually be driving the Waymo cars the majority of the time because a $.50/hour worker is cheaper than the AI compute costs for inference in a 4D driving model.

Without transparency about how these AI surveillance systems really work, there could be a room full of random men in a foreign country who spend their entire day watching the video feeds (and saving recordings) from college dorm hallways or elementary school classrooms.

Dangerous False Alarms

AI gun detection companies have contracts at k-12 schools and universities across the country despite a lack of 3rd party testing or documentation about the performance of the systems. Accuracy and false alarms are a primary reason William & Mary students voted to cancel the contract.

In December 2025, students at Lawton Chiles Middle School in Florida were sent scrambling into lockdown after an alert from the ‘ZeroEyes’ AI surveillance system detected a student carrying a gun. The “gun” was actually a band student’s clarinet.

In 2023, the same ZeroEyes AI software interpreted this CCTV image as a gun at Brazoswood High School in Clute, Texas, sending the school into lockdown and police racing to campus. The dark spot turned out to be a shadow on a drainage ditch that is lined up with a person walking. It’s likely that the human reviewers at ZeroEyes also misinterpreted this image because a person assigned to review images for guns is psychologically prime to see ambiguous objects as guns.

Another root cause of these errors is that cameras generate poor-quality images in low light, bright light, rain, snow, and fog. When the limitations of the camera systems are easy to see, should a school be using AI to make life-or-death decisions based on a dark, grainy image that an algorithm can’t accurately process?

William & Mary students weren’t the first ones to see these problems. A large transit system in Pennsylvania also canceled its multi-year contract with ZeroEyes after the initial pilot phase because the transit system said the software couldn’t reliably spot guns.

In 2025, AI CCTV monitoring systems not only dispatched police to false alarms for erroneous objects (e.g., band instrument, bag of chips), an AI system also failed to identify a real gun during an actual school shooting:

• AI system failed to identify a school shooting in progress at Antioch High

• AI classified Doritos chips as a gun and then police pointed real guns at students

• FTC take action on the false marketing claims of an AI metal detector company

• AI camera system dispatch police for an active shooter who was a band student holding a clarinet

• Another AI camera system dispatched an active shooter response for a student holding a bag of Doritos chips

• CCTV system at a major university failed to record photos of the assailant entering or leaving the campus after shooting 12 students inside a classroom

Errors by Design

Going back to sole-purpose image classification software from the 2010s, computer vision doesn’t see objects the way that humans do because AI software is based on math. Each pixel in the image is given a numeric value and then the program creates probability values that the proximity of clusters of pixels in training images are the same as a cluster of pixels in a new image.

While it sounds complicated, there are free open-source tools that can analyze how deep learning algorithms work and what they are “seeing” in the images.

Without any special training, an existing image classification model can recognize attributes of a toy gun and the model will assign probabilities that the toy is different types of real firearms.

With a picture of a real handgun, the same open-source model can classify this handgun as a firearm but isn’t really sure if it’s a revolver (incorrect), assault rifle (incorrect), or rifle (incorrect). But the exact type of gun doesn’t really matter if you are just trying to detect any type of firearm. The software’s confusion does become a problem if you are trying to determine the difference between a toy gel blaster versus a real gun (or a shiny Doritos chip bag with edges that make a 90 degree angle versus a shiny pistol slide).

Because a computer can’t see like a person, to understand how a model is classifying an image you need to calculate the gradients to measure the relationship between changes to a feature and changes in the model’s predictions. For comparing these two images of guns, the gradient shows which pixels have the strongest effect on the model’s predicted class probabilities.