“Where’s Daddy?” turned homecoming into a bombing opportunity, Israeli sources say.

The invention examined here is the conversion of traces into permission. A communications pattern is recorded. A system gives it significance. An institution decides how much confidence to place in the result, then decides what violence that result can support. Those are separate acts, even when an interface compresses the sequence into a few clicks.

AI is an umbrella term, not the name of one weapon. Machine learning can identify patterns and rank suspected affiliations. Other software tracks locations, organizes intelligence or stores intercepted material. Cloud storage is infrastructure; location tracking is not necessarily AI; a recommendation system is not automatically an autonomous weapon. Calling the entire chain “AI” makes the technology sound more unified, and responsibility more distant, than the evidence warrants.

In its November 2023 investigation, +972 and Local Call described Habsora, or “The Gospel,” as accelerating the production of potential structural targets. Its sources also alleged more permissive policies toward civilian harm. [2] Faster production and lawful selection are different measures. A system can increase the first without establishing the second.

Human Rights Watch’s September 2024 assessment distinguishes Lavender’s person ratings, The Gospel’s structural targets, location tracking and evacuation monitoring. HRW confirmed that the military used a tool with the location-notification functionality reported as “Where’s Daddy?”, while saying it could not confirm further technical details. It also warned that a phone’s location is an unreliable substitute for establishing a specific person’s presence. [3]

The evidentiary chain can break at several points. A device may be attributed to the wrong person. A pattern may resemble a militant’s activity for another reason. A location estimate may be stale. An assessment of who remains inside a building may be incomplete. These are possible failure modes, not findings about a named strike. Each requires its own check; a confidence score at one stage cannot certify every later stage.

The ICRC and Geneva Academy’s 2024 analysis describes automation bias: pressure and opaque outputs can encourage operators to accept machine recommendations instead of testing them. It also recognizes that carefully designed tools can help protect civilians. The decisive questions concern data quality, the task and the conditions under which people use the output. [9]

Meaningful review therefore needs evidence, time and authority to refuse. A human name on an approval record establishes participation. It does not, by itself, establish that anyone challenged the recommendation. The signature is where the examination should begin.

NAZA belongs in this case because it examines the people, policies and language behind the targeting process. The Venice festival record identifies an eighty-minute documentary directed by Yuval Abraham and Rachel Szor, produced by The Guardian and JW Films. In their September 2026 account, the directors say they interviewed twenty-four Israeli soldiers, mainly intelligence officers, and explain the title as a Hebrew acronym for “collateral damage.” [5, 6]

The directors argue that the testimonies describe a system of calculated civilian killing rather than isolated departures from policy. That is their stated conclusion, which the IDF contests. They also explain that some revelations were published earlier through +972, Local Call and The Guardian. The film and those investigations therefore have overlapping provenance. Repetition across a film and an article must not be counted as two independent witnesses. [6]

The language matters. “Collateral damage” is an administrative category applied to people whose lives continue beyond the target file: parents, children, neighbours. An estimate can acknowledge foreseeable deaths while keeping those people grammatically outside the main event. That is this article’s reading of the term, not a substitute for reconstructing any attack.

Viewing boundary: this article uses the festival record, the directors’ published account and the IDF response. The full film and its underlying interview recordings were not reviewed for this article. No scene-by-scene verification or independent authentication of its anonymous sources is claimed.

In its October 2026 response to NAZA, the IDF says AI neither determines target designation nor decides whether a strike occurs. It describes human analysts, authorized intelligence officers and a broader approval process culminating in an operational commander’s decision. It characterizes Lavender as an intelligence cross-referencing database, not an approved strike list, and disputes the interviewees’ understanding of the full process. [7]

The IDF also distinguishes thresholds determining which authority may approve a strike from permission to kill a fixed number of civilians. It says commanders must make individual proportionality assessments and that suspected violations are examined. Those are the military’s stated rules and account of its conduct. The sources’ allegations concern how the system operated in practice. A written procedure cannot alone settle a dispute about its implementation. [7]

HRW’s analysis supplies additional evidence but also limits: it examined apparently operational data exposed through an IDF evacuation website, while declining to claim definitive confirmation of that data’s origin and use. [3] Those limits belong in the record alongside the findings. A counter-record gains force through the precision of what it can establish.

The ICRC’s June 2026 guidance states that international humanitarian law applies to AI use: distinction, proportionality and precautions remain required, and humans remain responsible. [8] Whether a particular attack violated those duties requires the facts of that attack. “AI was involved” is neither proof of a war crime nor a defence against one.

The public record supports scrutiny of digital tools in Israeli intelligence and targeting, serious insider allegations about operational practices, and a documented restriction of specified Microsoft services. It also records explicit military denials. The public sources examined here do not provide a complete technical audit of Lavender or a strike-by-strike accounting of AI’s contribution to deaths.

Accountability would require preserving the relevant model version, input provenance, analyst assessments, location checks, civilian-harm estimates, approval decisions and later casualty reviews, with appropriate access for independent investigators. This is an evidentiary demand arising from the dispute, not a claim that those records have been disclosed.

The political question survives every technical qualification: who decided that the available evidence justified the attack? Someone selected the system, approved its use and set the conditions for acting on it. The presence of software adds steps to examine. It does not remove the people who made those choices.

Read with the Snowden case, the recurring problem is unequal visibility: institutions can inspect a population’s traces while keeping the instructions governing that inspection out of public view. In Gaza, examining that imbalance also means examining decisions about lethal force.