This article is from: srnnews.com
Behind NAZA: How AI Is Used to Assess and Reduce Collateral Damage
Cyber intelligence expert Matías Rott says the same systems scrutinized by NAZA are used to assess and reduce expected civilian harm before a strike
By Gabriel Colodro / The Media Line
The title of one of the most discussed films to emerge from this year’s Venice Film Festival is not the name of an artificial intelligence system. NAZA is Israeli military shorthand derived from nezek agavi: collateral damage. It refers to the unintended harm expected beyond the military target of an attack. That distinction is also at the center of a problem with the documentary’s framing. If the Israeli military is calculating expected civilian harm before an attack, what is that calculation intended to accomplish?
The 80-minute documentary by Israeli filmmakers Yuval Abraham and Rachel Szor premiered in Venice on September 10 and received the festival’s Special Jury Prize. The official Venice synopsis says the film examines what it describes as systems behind Israel’s “calculated mass killing” of Palestinian civilians in Gaza. Its case relies heavily on anonymous Israeli military and intelligence personnel describing targeting procedures, artificial intelligence tools and levels of civilian casualties allegedly accepted during the war.
Yet calculating collateral damage is not, by itself, evidence that civilians were selected as targets. The Israeli military says the assessment exists for the opposite reason: commanders are required to estimate the likely incidental harm from each strike, compare it with the anticipated military advantage and consider measures that could reduce civilian harm. Under the Israel Defense Forces’ (IDF) published procedures, a strike should not proceed if the expected collateral damage is excessive in relation to that advantage.
Cyber intelligence expert Matías Rott said the meaning of the term is crucial to understanding the argument. “NAZA means collateral damage,” Rott told The Media Line, describing harm caused to people or structures that were not themselves the intended objective of a strike.
For Rott, the use of artificial intelligence to calculate that harm is fundamentally a decision-support function. Such technology can integrate large volumes of intelligence, model a target’s surroundings, and help assess variables that may affect the consequences of an attack. “When you are working with this AI, for sure what you are willing or intending is to minimize collateral damages,” he said.
Rott gave the example of attacking a building from different directions. An impact on one side could place another structure in the blast area, while an attack from another direction could face an open space. The weapon selected, its precision, and the direction from which it is used can therefore change the expected harm outside the target.
“All these kinds of things are taken in consideration,” he said, with artificial intelligence helping assess how to conduct an attack “in order to minimize the collateral damage” while still attempting to achieve the military objective.
The dispute surrounding NAZA, then, is not really over whether Israel measures collateral damage. Both the military and the filmmakers’ own previous reporting establish that it does. The more consequential dispute is what limits were placed on that damage, how reliable the estimates were, and what happened when the predicted civilian toll approached those limits.
That distinction is visible even in Abraham’s reporting from two years before the film.
In April 2024, Abraham reported for +972 Magazine and Local Call that the military used a system known as Lavender to process intelligence concerning suspected Hamas and Palestinian Islamic Jihad operatives. Six intelligence sources alleged that officers relied heavily on its outputs during the early stages of the war, sometimes providing only limited human review.
But the same investigation repeatedly describes collateral-damage calculations as a condition for whether a strike could proceed.
One intelligence source quoted in the report said that after Lavender identified a suspected operative and a human checked the target, bombing remained “subject to an examination of collateral damage.” Elsewhere, the investigation says some attacks against suspected junior operatives were not carried out when the expected number of civilians in the building exceeded the permitted collateral-damage level.
That does not settle the controversy in Israel’s favor. The sources made a different and serious allegation: that during parts of the war the permitted threshold itself was too high.
According to the documentary, two sources told Abraham that during the initial weeks of fighting, attacks against some suspected low-ranking fighters could be authorized with an anticipated civilian toll of up to 15 or 20 people. The IDF disputed that description, saying that targets were individually assessed and that strikes were not conducted when expected collateral damage was excessive relative to the military advantage.
The same investigation also challenged the accuracy of the calculations. Sources said automated estimates of how many civilians remained inside buildings could rely on prewar residency figures adjusted to broader evacuation estimates, rather than on direct, real-time confirmation of who was present.
Those allegations point to a more precise criticism than simply saying that Israel “calculated” civilian casualties. A collateral-damage system can be intended to reduce civilian harm and still be criticized for inaccurate inputs, permissive thresholds, or insufficient human scrutiny. The existence of the calculation and the quality of that calculation are separate questions.
The IDF also rejects portrayals of Lavender and Habsora as autonomous targeting machines. In its public explanation of the systems, the military describes Habsora as a tool that brings potentially relevant objects to an intelligence analyst’s attention and Lavender as a database that cross-references existing intelligence about suspected operatives. Neither, the military says, can independently establish that a person or object is a lawful military target.
Rott similarly said artificial intelligence can dramatically accelerate analysis but does not eliminate the human chain of command. “For sure, the last call is going to be a human being,” he said. “Somebody has to accept, somebody has to push a button, somebody has to agree in order to perform an attack.”
He nevertheless identified a legitimate question raised by the film: whether that person is independently evaluating the intelligence or merely approving an output produced by automated systems. “The dilemma right now,” Rott said, is whether the human was acting “as a rubber stamp” or genuinely processing the information before making the decision.
The IDF has categorically rejected NAZA’s allegations and challenged the filmmakers’ reliance on anonymous witnesses whose identities, positions and access to the decisions they describe cannot be publicly verified. Chief of Staff Lt. Gen. Eyal Zamir has also ordered further action over the film, including an examination of whether classified material was leaked.
What the filmmakers ultimately establish will be easier to assess once NAZA is released more widely. For now, the terminology itself matters. A military does not need a collateral-damage estimate to know how effectively it can destroy a target. It needs that estimate to know what else may be destroyed with it, and to decide whether and how the strike should proceed. The dispute left by NAZA is therefore not why Israel calculated civilian harm, but whether the limits placed on that harm were adequate and whether the calculations on which those decisions rested could be trusted.
Brought to you by www.srnnews.com
