The searing
emotion of this finale is a sharp tonal break from NAZA’s preceding
scenes, which are composed of nighttime interviews with 24 Israeli intelligence
officers. The core of each interview is an often quite clinical description of
the systems these soldiers used—and may be still using—to target and kill
Palestinians. Though of course the Israeli government has gone on offense to label NAZA “distorted, misleading,
and disingenuous”—with Prime Minister Benjamin Netanyahu threatening to revoke
the directors’ Israeli citizenship—the film is remarkably straight in its
presentation. Abraham, who conducts each interview, is shown on-screen; we hear
his questions and how he sometimes pushes back on his interviewees’ answers.
We also see how an interview was interrupted by an air raid siren, a regular
fixture of life for Israelis. Though the cinematography is effective, there is
no filmmaking magic or trickery here, no moments that a reasonable person could
describe as propaganda. Just 24 Israelis explaining from their own firsthand
experience how they helped kill thousands of unarmed men, women, and children
in Gaza.

Much of NAZA’s
run time focuses on descriptions of two related AI surveillance and targeting
systems used by many of the interviewees. The first, named Lavender, identifies Hamas operatives—or,
more accurately, the phones of Hamas operatives—through their behavior. So if
someone calls another person whom the system has identified as related to
Hamas, they’re awarded a certain number of points; if they go to a location the
system has identified as related to Hamas, they get more points. Once the
person has enough points, they’re automatically added to Israel’s kill list.
“It’s a perfect job for machine learning,” noted one officer who helped develop
the program. “Looking
at characteristics, making a determination. Is it a cat or a dog?” Lavender allowed Israel to unleash “targeted” attacks at
scale. “Efficiently, like in a factory,” said an interviewee—attacks directed
not just against senior Hamas officials but people “so junior we didn’t know
who they were.” It was, said another officer, about “quantity, not quality.” Of
course, as anyone who’s encountered a large language model hallucination knows, AI regularly
makes mistakes. And this was factored in too. As one officer explained, Israel
operated with an understanding that its Lavender kill list had a 10 percent–15 percent
error rate. 

The second
system, nicknamed  “Where’s Daddy?”
identifies the right time to bomb a target. The right location is already
decided—Israel’s protocol, the officers explain, is to strike targets in their
homes. To determine whether the person is actually at their home, Where’s Daddy? monitors their family (again through their cell phones), searching in
particular for children asking their mothers when their fathers would be
home—hence the program’s name. One officer described monitoring a 12-year-old
girl and then calling in a strike, saying, “You hack her phone and then you
kill her.” Another remembered hearing a baby crying in the background before he
signaled for an attack. In a third conversation, an officer described looking
through the photographs saved on a target’s phone to help identify the best
angle of attack on the man’s apartment. He recounted seeing a picture of the
target with his young daughter, then listening in real time to the family’s
screams as the girl was burned alive.