Yesterday, Lawton Chiles Middle in Oviedo, FL (just north of University of Central Florida and ~20 minutes from my house) went on 'CODE RED LOCKDOWN' for a person with a rifle on campus.
But it wasn't a rifle, the AI gun detection system flagged a clarinet--the wooden musical instrument that's used in every marching band--as a gun. The system has military veteran human reviewers and their assessment of the image also determined the clarinet was a gun and they dispatched police to the campus. (Below is the video clip about the vendor the Seminole County School District signed a contract with in 2024)
This combined software and human failure was highly predictable and has happened at multiple campuses this fall because:
CCTV images don't generate enough data to accurately classify small objects like guns objects (edge of a clarinet or a Doritos bag don't have enough pixels to be different from a handgun)
Human reviewers who are employed to identify images of guns are psychologically primed to see ambiguous objects as guns. When they are forced to make a decision within 2-3 seconds (per the video from the vendor), decades of psych studies show our 'System 1 fast response' thinking is prone to errors.
Image data alone can't provide reliable computer vision in complex environments like crowded hallways and dark areas outside the school (this same system sent an active shooter alert for a shadow). This is why Waymo self-driving cars use multiple cameras plus LiDAR to create a 4D picture of objects and depths.
Last year at a board meeting about the AI gun detection contract, a parent said: “I would like to see some statistical data that proves that it does work, and it works. In the time and years Seminole County has had it, has it worked? How many people have they stopped? And as parents, we should know these things. If they are happening, why are they not being reported?”