The AI towers watched the border and nobody came

MIT Technology Review published an investigation this week, produced with Times of San Diego under the title "Dying on Camera", into deaths along the US-Mexico border's "virtual wall" of AI-enabled surveillance towers. The reporting opens with José Morales Bernal, who crossed into the United States on 8 April 2024, the day before his 32nd birthday, and walked through southern New Mexico within range of three surveillance towers. The newsroom spent the past year building the first comprehensive map of deaths near those towers, and its finding is that people walked undetected through surveilled ground, died nearby, and lay unnoticed. The government spent billions on the system. The system watched.

Detection is not response

The consensus reading of a story like this is procurement failure: the wrong vendor, the wrong cameras, a budget that outran the engineering. That reading is comfortable because it implies the fix is a better tower.

The actual failure sits one layer up. A detection system only produces value when a named human is accountable for acting on what it detects, inside a defined window, with consequences when they do not. Coverage was installed. The chain from alert to human response was not.

This is the same structure now being installed inside ordinary organisations. Monitoring tools that flag anomalies nobody owns. Risk models that score accounts nobody calls. Copilots that draft the recommendation while the person who used to make it is reclassified as a reviewer. In each case the sensing layer gets funded and the deciding layer gets thinned, because sensing is what vendors sell and deciding is what headcount costs.

The border is the extreme version because the cost of the gap is counted in bodies. The mechanism is not extreme at all. It is the default outcome when you automate perception and leave accountability implicit.

What this tells an operator about their own position

Two things.

First, the work that survives automation is rarely the work that produces the signal. It is the work that owns the consequence of the signal. In the border system, the towers did their job in the narrow sense — they were built, deployed, aimed. The unowned part was the response. Anyone whose role is generating outputs that someone else is responsible for acting on is sitting in the absorbable layer.

Second, systems like this create a specific illusion for the people inside them. Coverage reads as control. When a dashboard is green and a tower is live, the organisation believes the problem is handled, and the humans who would have noticed otherwise stop looking. MIT Technology Review needed a year and a second newsroom to establish what the deployed system did not surface on its own. That gap between what a system observes and what anyone does about it is where careers, and in this case lives, are lost.

If you are being asked to justify your role against an AI tool this year, the argument is not that you see things the tool cannot. Increasingly it will. The argument is that you are the person who carries the outcome.

Your Next Move

Audit one system you rely on for its response chain, not its accuracy. Pick the AI or monitoring tool closest to your work. Write down what happens between the moment it flags something and the moment a named person acts. If you cannot name the person and the time window, the system is decorative and you can say so in a meeting this week.

Move your job description from output to consequence. List what you produced last quarter. Next to each item, write who was accountable if it was wrong. Where the answer is not you, that is the work most exposed to absorption. Negotiate to own the decision, not just the deliverable.

Keep one manual check that the automation does not touch. A weekly sample you inspect yourself. It is how you stay the person who notices when the green dashboard is lying.