‘Lavender’: The AI machine directing Israel’s bombing spree in Gaza

🌈 Abstract

The article investigates the Israeli army's use of an AI-based program called "Lavender" to automatically generate thousands of targets for assassination during the recent war on the Gaza Strip. It reveals how the army relied heavily on this system to mark low-ranking Palestinian militants and their homes for bombing, often with little human oversight and high civilian casualties. The article also discusses other automated systems used to track and locate these targets in their homes, as well as the army's loosening of restrictions on collateral damage.

🙋 Q&A

[01] Generating Targets

1. What is the Lavender system and how did it work?

• Lavender is an AI-based program developed by the Israeli army to automatically mark tens of thousands of Palestinians, including low-ranking militants, as potential targets for assassination.
• Lavender analyzes data collected through mass surveillance to assess the likelihood of each person being affiliated with the military wings of Hamas or Palestinian Islamic Jihad (PIJ). It gives each person a rating from 1 to 100 based on various "features" that indicate potential militant activity.
• The army relied heavily on Lavender's assessments, with officers only briefly checking that the target was male before approving strikes, despite knowing the system had an error rate of around 10%.

2. Why did the army develop Lavender?

• Prior to the war, the army only tracked and targeted senior Hamas commanders, which required a lengthy "incrimination" process.
• After the October 7 attacks by Hamas, the army decided to designate all Hamas operatives, including low-ranking ones, as legitimate targets.
• This posed a technical challenge, as manually processing the large number of potential targets was seen as a "human bottleneck". The solution was to automate the process using Lavender.

3. What were the consequences of the army's reliance on Lavender?

• Lavender marked around 37,000 Palestinians as potential targets, most of them low-ranking militants or civilians with loose connections to Hamas.
• The army gave sweeping approval to adopt Lavender's lists, with little human verification of the system's assessments.
• This led to the bombing of thousands of Palestinians, most of them women, children, and uninvolved civilians, in the first weeks of the war.

[02] Linking Targets to Family Homes

1. How did the army identify the locations of the Lavender-marked targets?

• The army used additional automated systems, such as "Where's Daddy?", to track the targets and alert when they entered their family homes.
• By linking the Lavender-marked individuals to their family residences, the army could easily locate and bomb them, even if they were not engaged in any military activity.

2. What was the impact of this approach on civilian casualties?

• Bombing targets in their homes, rather than during military activity, resulted in the deaths of thousands of civilians, often entire families.
• More than half of the fatalities in the first month were from 1,340 families that were wiped out in their homes.

[03] Choosing a Weapon

1. What type of munitions did the army use to strike the Lavender-marked targets?

• For junior operatives marked by Lavender, the army preferred to use "dumb" unguided bombs, which are cheaper but cause more collateral damage, rather than more precise "smart" bombs.
• This was done to save the more expensive guided munitions for higher-value targets.

[04] Authorizing Civilian Casualties

1. What were the rules of engagement regarding civilian casualties?

• The army significantly loosened the restrictions on the number of civilians that could be killed as "collateral damage" when targeting junior militants.
• In the initial weeks, the army authorized the killing of up to 15-20 civilians per junior operative, a much higher number than in previous wars.
• For senior Hamas commanders, the army at times authorized the killing of over 100 civilians in a single strike.

2. What was the rationale behind these permissive rules of engagement?

• The sources suggest the army was driven by a desire for "lethality" and a "vindictive" atmosphere after the October 7 attacks by Hamas.
• There was also pressure to generate as many targets as possible, leading to a loosening of restrictions on collateral damage.

[05] Calculating Collateral Damage

1. How did the army estimate the number of civilians in targeted buildings?

• The army used automated systems to estimate the number of residents in each building, based on pre-war data and assumptions about evacuation rates.
• However, these estimates were often inaccurate, as the systems did not verify the actual number of people present during the strikes.

2. What were the consequences of this automated approach?

• The inaccurate collateral damage estimates led to the bombing of buildings with more civilians than expected, resulting in higher civilian casualties.
• The army did not conduct thorough post-strike assessments to determine the actual civilian toll, instead quickly moving on to the next target.

[06] Bombing Family Homes

1. What happened when the army's intelligence on the target's location was inaccurate?

• There were instances where the army bombed a house based on intelligence that the target was present, but the target had actually left the house earlier.
• This resulted in the deaths of entire families who were not the intended targets.
• The army did not always verify the target's location in real-time before carrying out the strike.

2. How did this impact the army's post-strike assessments?

• In contrast to previous wars, the army did not conduct thorough post-strike assessments to determine the actual civilian casualties for junior militants targeted by Lavender.
• This lack of accountability meant the army did not know the full extent of civilian deaths resulting from these strikes.