Product1 publisher3 min readPublished
MIT Technology Review asks CBP for the tower audit it could have run since 2015
Reporters covered nearly 600 of the 800 towers the government says are operating, and they had to assemble the map themselves from outside sources. Death records follow no shared protocol, so their count of more than 1,050 deaths nearby is a floor.
The Product Desk · Product desk
What happened
- MIT Technology Review found cases of people who walked undetected through areas watched by AI-enabled border towers and later died nearby, their bodies unnoticed for weeks or months.
- The publication attributes those deaths to three flaws in the network: broken towers, algorithms that failed to detect people, and agents who did not respond to alerts.
- CBP publishes no comprehensive map of the virtual wall, so the reporters assembled tower locations and approximate install dates from outside mapping work and satellite imagery.
- CBP has been able to audit tower effectiveness since at least 2015, when it started installing this generation of towers, and sources told the publication it has never done one.
- The US is set to spend $1 billion to triple the size of the virtual wall by 2034.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint With the tower inventory incomplete outside CBP and the death records scattered across agencies, nobody can compute a detection rate for the system, only a floor for how many people died within range of it.
- decision The next tranche of towers is being sized on deployment counts. The effectiveness measurement that would argue for or against buying more of them has not been produced.
- contradiction Two of the three reported failure modes are the hardware and the model, the part vendors warrant.
- precedent An agency that held the install and uptime records for a decade and never set them against outcomes sets the working standard other AI detection buyers will be held to.
An alert lands on a screen in a Border Patrol station and nobody drives out to look. MIT Technology Review and Times of San Diego put that at one end of the chain in Dying on Camera, their year-long examination of the border surveillance network [2]. At the other end are a tower that was not working and a model that failed to detect the person walking past it [4]. Two of the three failures are equipment and software [4].
To say whether a detection system works you need two registers: where the sensors were and when they were live, and what happened to the people they were supposed to see. Neither is public. Working from Electronic Frontier Foundation mapping and satellite imagery, the reporters covered nearly 600 towers against the 800 the government says operate today [6][7]. Roughly 200 units stayed outside the sample, about a quarter of the fleet [1]. Install dates are approximations. The two newsrooms call their count of more than 1,050 deaths near towers an undercount [8].
The death records are worse. No national agency collects and publishes all of them [12]. In Texas alone the reporters went through thousands of pages from dozens of agencies and still describe their picture as incomplete; each agency documents these deaths in public records its own way [13].
Border Patrol has run its own tracking system, BSITS, since 2007, and national rules for how sectors should enter cases into it arrived in 2021 [14][15]. For fourteen years, each sector decided for itself [2]. The GAO found in 2022 that many cases still were not in the system, particularly deaths Border Patrol had not discovered itself [16]. In 2023 it reported that steps had been taken, and outside researchers still put the agency's figures below the true number [17]. CBP began installing this generation of towers in 2015 [10], and ten years later, in April 2025, the GAO's most recent look at the prevention side found the Missing Migrant Program improving how it collected data on events like 911 calls and rescues [18][3].
The first of the four fixes MIT Technology Review proposes is a comprehensive audit of deaths near the virtual wall [19]. Shown that more than 18 people had died in 2024 near towers in one small stretch of desert near her district [20], New Mexico state representative Sarah Silva said: "If I were the federal government, I would do an audit, because this feels like it's not working." [21]
Any detection deployment can keep three counters, and they are not equally cheap. Uptime and detection volume come out of the system itself, nightly, at no extra cost. The third one, what happened to the people you were supposed to find, sits in other agencies' records. Reporting this system meant reading through dozens of county and state agencies [13]. Put to a fleet expansion, the useful question is which of the three counters more units will move. Anduril, which built the latest towers, continues to win major contracts and is positioned for the federal border security money [22].
What to watch
- Whether CBP publishes a tower map or an effectiveness audit covering installs since 2015.
- Whether the $1 billion expansion budgets outcome tracking alongside the new towers.
- Whether the GAO revisits BSITS undercounts after its 2023 finding that steps had been taken.