Inside Gaza's AI targeting: Lavender, Gospel, 'Where's Daddy?'

The interesting part of the Wikipedia entry collating what is publicly known about Israel's AI-assisted targeting in Gaza is not the model architectures. It is what the words 'human in the loop' actually meant when the loop was running at wartime tempo.

The entry pulls together reporting from +972 Magazine and Local Call, alongside The Guardian and others, on three named systems. The Gospel, developed by Unit 8200 of the Israeli Intelligence Corps, produced bombing targets from surveillance data. Retired Lt Gen Aviv Kohavi, head of the IDF until 2023, is quoted saying the system could produce 100 bombing targets in Gaza a day, where in the past the force would produce 50 targets in Gaza per year. Lavender was a separate database that at one stage listed as many as 37,000 Palestinian men linked by the AI system to Hamas or Palestinian Islamic Jihad, focused on 'junior' operatives; the unit concluded it had achieved a 90% accuracy rate. A third program, 'Where's Daddy?', tracked suspected militants until they returned home.

The bit that should make anyone selling or buying 'human review' language uncomfortable is what an intelligence officer told the reporters: 'I would invest 20 seconds for each target at this stage, and do dozens of them every day.' Another said they had 'zero added-value as a human, apart from being a stamp of approval'. The article cites pre-authorized civilian casualty ceilings that reportedly reached as low as five and as high as 20 uninvolved civilians per junior operative, and over 100 for top-ranking Hamas officials. Suspected operatives were, in one testimony, 'bombed in homes without hesitation, as a first option'.

The honest caveat is that the underlying testimony comes from intelligence officers speaking to +972 and Local Call, and the 90% accuracy figure is self-reported by the same unit that deployed the system. The Wikipedia page is itself a collation, so treat the specifics as reported rather than settled. UN Secretary-General António Guterres said he was 'deeply troubled' by reports that Israel used artificial intelligence in this way.

What the reporting doesn't give you is a ground-truth false-positive rate, the identity of any commercial components underneath these systems, or whether the workflow has been copied by other militaries. It is, however, the most detailed public case study we have of AI-generated kill lists at operational tempo, and any defense or enterprise buyer whose contract leans on 'a human will review this' should probably ask, out loud, for how long.