Product1 publisher2 min readPublished
Landfill workers found the man three border towers were supposed to see
MIT Technology Review's 15-month investigation into US border surveillance puts the failure downstream of the cameras, at the step where a detection was supposed to turn into an agent arriving.
The Product Desk · Product desk
What happened
- Jose Morales Bernal crossed into the US in April 2024, the day before his 32nd birthday, within range of three surveillance towers whose cameras and AI were built to detect and track people automatically.
- He died 360 feet from the closest tower, and local landfill workers rather than Border Patrol were the first to spot him; an autopsy gave the cause as "environmental exposure."
- MIT Technology Review spent 15 months mapping nearly 4,000 locations where human remains were found against information on nearly 600 border surveillance towers.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint More towers add more detections to the step that failed. The gap described in the reporting sits between an alert and an agent arriving, and a larger sensor network does not shorten it.
- contradiction Nominal range and likely field of view produce different coverage counts from the same towers, so an acceptance test written on datasheet range credits the system for ground it cannot see.
- cost The cost of an unanswered alert is borne by the person in the desert, not by the vendor or the agency, so the system generates no internal signal that it failed.
- precedent Buyers of alert-based detection AI now have a public case where the sensing goal and the stated outcome were measured separately and came apart. That case makes detection-rate-only acceptance criteria hard to defend.
A detection system and a response system sit inside the same procurement, and MIT Technology Review's reporting describes failures in each. Broken towers and algorithms that failed to detect people are the vendor's half. Agents who did not respond to alerts are the operator's half [9]. Improving the model does not touch the second one.
MIT Technology Review did not report whether any of the three towers registered Jose Morales Bernal. It did report what the design promised after a detection: agents were trained to provide medical help if the person needed it [2].
The two coverage counts in the reporting measure different things. Times of San Diego counted at least 138 cases since 2022 where remains were found within the nominal range of a nearby tower [11]. MIT Technology Review's own analysis estimated that more than half of those people were likely within a tower's field of view [12], which puts the lower bound above 69 of the 138 [15]. Nominal range is the figure on the datasheet. Field of view is what one camera sees once terrain, buildings and mounting angle are accounted for.
The next purchase is sized in towers. The US government is about to spend another $1 billion to triple the virtual wall [8], and nearly 600 towers tripled means roughly 1,200 new ones, which works out near $833,000 each [14].
Teams that ship alert products tell themselves the queue gets worked. What people on shift actually do is work the alerts they are measured on, and the rest age. The numbers that describe this system are not detections per tower or camera uptime. They are the share of alerts that ended in contact with a person, and how long that took.
For anyone rolling out detection AI on Monday, the acceptance test needs two axes: did the system produce the signal, and did a human act on it inside the window that matters. Signal plus action is the working case. If the signal arrives and nobody acts, that is a staffing and workflow defect, and retraining will not move it. The two kinds of miss split the same way. A miss where someone would have responded belongs to the model; a miss with nobody free to respond means the model is not the purchase to argue about. MIT Technology Review published four ideas for what Customs and Border Protection should do about the failures it found [10], as part of a series it calls Dying on Camera [13].
What to watch
- Whether CBP publishes alert-to-contact times or response rates for the towers already standing.
- How many towers the $1 billion actually adds, since tripling is a nominal figure and site costs vary.
- Any CBP response to the investigation, including whether it disputes the mapping of remains locations against towers.