AI in Gaza: Siri or Skynet?

A concerning article was recently published in the left-wing Israeli magazine +972. Drawing on the testimony of six anonymous Israeli intelligence officers, the article paints a dark picture of Israel’s use of AI technology. Specifically, they referenced a targeting program called “Lavender” and a tracking program uncomfortably named “Where’s Daddy”.

The integration of  AI in military technology has become an inevitable development. AI’s capabilities in computation and identification can enhance the precision of targeting enemy combatants, thereby increasing effectiveness and reducing civilian casualties. Despite evoking dystopian concerns, AI has already been incorporated into advanced military systems globally. Some complex tracking and data collection systems have been deployed as early as 2007 in Iraq but the advent of machine learning has taken these systems to the next level.

The Lavender system supposedly functions by identifying Hamas activities and through surveillance, intelligence, and tracking data, determines the likelihood of someone being a Hamas militant. The “Where’s Daddy” system supposedly ties them to locations and the likelihood of being there at certain times. The system supposedly generated 37,000 targets in the early stages of the war, many of them tied to family homes with “Where’s Daddy” tracking data, leading to high civilian casualties. However, the content or nature of this list is unknown and this list could simply be those with some chance of being affiliated.

Israel has always stated with pride that airstrikes undergo several checking stages before approval. However, the testimony of some of these officers describes those checks as a rubber stamp with only 20 seconds per strike, with reviewers only determining if the target was male. This is despite a supposed 10% margin of error for the Lavender system.

One of these officers claimed that the implementation of “Where’s Daddy” led to a systematic targeting of homes of often junior militants with significant family casualties. Additionally, one of the sources claims that acceptable collateral damage for strikes on more junior militants was as high as 15-20 and as high as 100 for one Hamas officer struck in November. Another testimony claims that the choice of munition was often excessive, using large unguided bombs for unnecessary targets. Though the source claims that these policies only applied in the first weeks of the war.

The IDF officially denies the claims of disproportionality and recklessness, maintaining that strikes go through sufficient checks before approval and munitions are selected carefully for the target. They also claim that Lavender is used only as one component of Israel’s intelligence gathering to cross-reference separate databases regarding potential targets.

There is confirmation of Israel’s use of the “Gospel” system, which identifies all manner of Hamas targets including infrastructure, launch sites as well as operatives. The IDF argues that the system assists operations in finding active Hamas targets which are used as recommendations for personnel. The accusation is still one that is deeply concerning to a defender of the IDF in Gaza.

There are reasons to be skeptical about the accusations and the framing of the information. As previously mentioned, systems used to locate targets and their homes have been used by the US and other militaries in their own operations. The focus on the AI system is intended to evoke dystopian imagery and cast Israel as some sort of sci-fi villain.

On its own, the use of such systems is in no way incriminating so long as the correct protocols are followed. In December the Gospel system was subject to a freedom of information appeal which confirmed that the reasons for the Gospel’s decision-making are provided to the officer along with its recommendation.

While the fact that Gospel generates more potential targets than previous methods is framed as excessive, ultimately it is intended to ensure the safety and success of the IDF, the more potential targets that may have previously gone unnoticed could be considered a success rather than a condemnation.

There is also no necessary evidence of disproportionality in the initial casualties of Israel’s bombing campaign. There is always a higher likelihood of higher casualties in initial bombings as the evacuations were still ongoing and ground forces had not moved in yet. Casualties are usually at the highest during this initial period as militaries prepare for maneuver of ground forces and target the major centers of operations and infrastructure.

By November 1^st the IAF conducted more than 11,000 strikes, including “anti-tank launch sites, Hamas commanders, and other threats.” It is unlikely that all of those were targets identified by Lavender or other systems and that none had proper checks before implementation.

Total casualties, according to the Hamas-run Gazan Ministry of Health, at 8,800, around 6,000 coming from 800 families. Israel claimed around 5,000 militants were killed by December as well as the destruction of large amounts of weapons and infrastructure.

Given the recent revelations about how the Gazan Health Ministry collects data, we can cast some doubt on this number, especially given the falsification of higher casualties of women and children. While this makes it unlikely that 20 or more civilian casualties were the norm for strikes, this does not fully absolve Israel.

As to the prevalence of casualties within single families Gazans tend to live in extended family homes meaning the bombing of a single building is more likely to have higher casualties from a single-family unit. Additionally, clan structures within the territory may mean that distant relations are reported as family members. There is also a higher likelihood that if one member of a family is involved with terrorism others will as well, peer influence is a significant factor in radicalization.

As to the accusation of munitions, as of Dec 21^st, only 208 of the most destructive 2,000-pound bombs were used which include munitions designed for breaching tunnels. It is unknown whether guidance systems were attached or what they were targeting.

The amount of destroyed or damaged buildings in Gaza is approximately 57% percent overall and 75% in Gaza City where most of the initial bombing was conducted. This is roughly similar to the 80% of destroyed or damaged homes in the high-intensity urban combat zones of Mosul and the 60% overall of Falluja. This indicates that Israel may not have used excessive amounts of munitions.

Israel’s more limited resources have made the use of smart bombs more costly compared to the US leading to 55-60% of munitions being unguided. That does not make the “dumb bombs” indiscriminate. Targeting systems are still used by pilots to calculate where to release munitions. Additionally, the 55-60% statistic was released in mid-December, after operations had begun in the north and the use of M117 demolition bombs for evacuated areas to create a buffer zone was documented.

Ultimately while we can point out reasons that Israel’s early bombing campaign does not seem to have been disproportionate, the truth of the strikes and the exact nature and use of Lavender and other systems cannot be assessed until more details are published.