Evidence track inside a parent dossier
claim-2026-lavender-automated-civilian-kill-lists
Overall verdict
Israel's use of AI systems like “Lavender” means machines choose civilians for death.
After April 3, 2024, +972 Magazine/Local Call reporting on “Lavender” and related tools (e.g., “Where’s Daddy?”, “Gospel/Habsora”) spread widely via mainstream outlets. The strongest version of the claim asserts that AI autonomously generated kill lists and that human review was a rubber stamp, so machines effectively selected civilians to die.
What is well-evidenced: Israel uses AI-enabled systems (e.g., Habsora/“Gospel”) to generate target suggestions and to fuse intelligence at scale. Investigative reporting based on unnamed Israeli officers alleges that “Lavender” mass-marked suspected militants, human checks were cursory, and strikes often hit homes at night—outcomes that correlate with high civilian harm. What is not shown by primary official records: that a machine, without human authorization, autonomously selected civilians as such for death. The IDF states analysts and commanders must verify and approve targets and denies using an AI that identifies ‘whether a person is a terrorist.’ UN reporting notes the allegations but also records the IDF’s position and says it could not independently verify the “Lavender” account. Bottom line: AI-assisted target nomination likely influenced pace and scale of strikes; claiming that ‘machines choose civilians for death’ overstates the public record and collapses disputed practice (minimal human review, high tolerated collateral harm) into autonomy and civilian intent. Label: misleading (not false), pending further verifiable primary evidence on autonomy and human oversight. Round 15 candidate-promote-high book gate (2026-06-13): High-confidence misleading: the record supports a serious AI-targeting-methodology question, not the public shortcut that machines autonomously generated civilian kill lists. The claim now has enough public counter/method/legal record to be high-confidence misleading. Guardian/+972/CNN preserve the allegation chain; Lieber Institute, JINSA, Modern War Institute, Washington Institute and IDF-targeting context prevent the public claim from becoming proof of automated unlawful killing.
If machines autonomously select human targets, that would upend core targeting rules (distinction, proportionality, precautions) and heighten risks of unlawful civilian harm and accountability gaps.
This page tests one narrow factual, legal, source-chain, or LOAC component inside a broader dossier.
Court, official, military/LOAC, watchdog, or explicitly role-labeled high-value material.
Context, methodology, legal analysis, and assessment-supporting sources.
Videos, transcripts, debates, timestamps, or source pages that prove what was said or published.
Allegation and amplification records; useful for tracing the claim, not proof of the accusation.
This file has explicit source-chain edges; read the sequence below before treating repetitions as independent proof.
Rotate, zoom, and select nodes to see how the claim and its evidence sources sit together. Click a node to zoom into it; double-click a claim or evidence node to open it. This is the exploratory view; the source list below remains the audit view.
Sources allege “Lavender” marked tens of thousands for assassination; outputs were treated “as if it were a human decision,” with ~20-second checks.
Primary investigative account alleging machine-generated kill lists and cursory human checks; lead requires verification.
Open sourcehttps://www.972mag.com/lavender-ai-israeli-army-gaza/
Article reports Lavender ‘achieved a 90% accuracy rate,’ prompting sweeping use to identify human targets.
Mainstream amplification of +972/Local Call findings; includes IDF response; lead for verification.
Open sourcehttps://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes
Mainstream amplification of +972/Local Call findings; includes IDF response; lead for verification. Bottom-source disposition: preserves allegation/transmission only; not bottom-source proof. The locator preserves where the allegation appears, not that the allegation is proven. Treat this as source-chain provenance unless separate primary records and counter-audit evidence satisfy the factual and legal threshold.
Open sourcehttps://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes
Primary investigative account alleging machine-generated kill lists and cursory human checks; lead requires verification. Bottom-source disposition: preserves allegation/transmission only; not bottom-source proof. The locator preserves where the allegation appears, not that the allegation is proven. Treat this as source-chain provenance unless separate primary records and counter-audit evidence satisfy the factual and legal threshold.
Open sourcehttps://www.972mag.com/lavender-ai-israeli-army-gaza/
Military and legal expert report on the October 7 war, Gaza operational context, Hamas strategy, civilian-harm mitigation, and LOAC framing. Matched by Priority-A source family: loac, intent, aid.
Open sourcehttps://jinsa.org/wp-content/uploads/2024/05/JINSA-Report-The-October-7-War.pdf
Background on AI target-bank; cites IDF material about producing targets ‘at a fast pace.’ Bottom-source disposition: preserves allegation/transmission only; not bottom-source proof. The locator preserves where the allegation appears, not that the allegation is proven. Treat this as source-chain provenance unless separate primary records and counter-audit evidence satisfy the factual and legal threshold.
Open sourcehttps://www.theguardian.com/world/2023/dec/01/the-gospel-how-israel-uses-ai-to-select-bombing-targets
Urban/subterranean warfare source for Hamas tunnel strategy, embedding, command infrastructure, and military-objective context. Matched by Priority-A source family: loac, intent.
Open sourcehttps://mwi.westpoint.edu/gazas-underground-hamass-entire-politico-military-strategy-rests-on-its-tunnels/
High-confidence misleading: the record supports a serious AI-targeting-methodology question, not the public shortcut that machines autonomously generated civilian kill lists.
Locator: Round 15 pass over candidate_promote_high claims from the 2026-06-13 medium/unclear triage.
Quote rule: Use the report summary and claim table; not an external factual authority.
https://worldagainstisrael.org/admin/books#round15-candidate-promote-high-book-gate
The claim now has enough public counter/method/legal record to be high-confidence misleading. Guardian/+972/CNN preserve the allegation chain; Lieber Institute, JINSA, Modern War Institute, Washington Institute and IDF-targeting context prevent the public claim from becoming proof of automated unlawful killing.
Locator: Round 15 pass over candidate_promote_high claims from the 2026-06-13 medium/unclear triage.
Quote rule: Use the report summary and claim table; not an external factual authority.
https://worldagainstisrael.org/admin/books#round15-candidate-promote-high-book-gate
Carries IDF’s formal denial that AI identifies terrorists and states analysts must verify targets.
Open sourcehttps://amp.cnn.com/cnn/2024/04/03/middleeast/israel-gaza-artificial-intelligence-bombing-intl
Methodology source for UN casualty reporting, source-chain attribution, and demographic/civilian inference limits. Matched by Priority-A source family: casualty.
Open sourcehttps://www.washingtoninstitute.org/policy-analysis/untangling-uns-gaza-fatality-data
High-authority LOAC methodology source for IDF targeting process, legal-adviser involvement, distinction, proportionality, and precautions. Matched by Priority-A source family: loac.
Open sourcehttps://www.lawandisrael.org/wp-content/uploads/Topics/Gaza/2023-Conflict/Lieber/lieber.westpoint.edu-Israel-Hamas-2023-Symposium-Inside-IDF-Targeting.pdf
LOAC source for why conduct-of-hostilities assessment in Gaza requires ex-ante, incident-specific evidence rather than effects-only inference. Matched by Priority-A source family: loac.
Open sourcehttps://www.lawandisrael.org/wp-content/uploads/Topics/Gaza/2023-Conflict/Lieber/lieber.westpoint.edu-Israel-Hamas-2023-Symposium-Assessing-the-Conduct-of-Hostilities-in-Gaza-Difficulties-and-Possible.pdf
This claim is already classified as misleading. The hostile source-chain remains preserved so readers can see where the allegation travelled. It does not prove the public overclaim unless the original bottom evidence independently proves the allegation and survives counter-audit. Current audit counts: 6 claim-side item(s), 4 allegation-chain item(s), 1 primary-like item(s), 17 counter/primary item(s).
Locator: P2 debunked-source-chain batch after archive-wide bottom-source proof audit.
https://worldagainstisrael.org/method#p2-debunked-source-chain-layer
Medium/unclear book-gate category: candidate_promote_high. This claim is public archive material, but print use depends on source-role caveats, counter-record weight, and pro-Israel book strength.
Locator: 2026-06-13 pass over all medium/null confidence claims after Round 14 review queue closure.
Quote rule: Use the report summary and claim table; do not quote as an external factual authority.
https://worldagainstisrael.org/admin/books#medium-unclear-book-gate-triage
Military context for ground operations in Gaza, tunnel/urban constraints, and operational factors absent from effects-only accusations. Matched by Priority-A source family: loac.
Open sourcehttps://mwi.westpoint.edu/these-are-the-challenges-awaiting-israeli-ground-forces-in-gaza/
Expert legal analysis summarizing allegations and IDF responses; discusses human-in-the-loop and LOAC framing. Bottom-source disposition: preserves allegation/transmission only; not bottom-source proof. The locator preserves where the allegation appears, not that the allegation is proven. Treat this as source-chain provenance unless separate primary records and counter-audit evidence satisfy the factual and legal threshold.
Open sourcehttps://lieber.westpoint.edu/gospel-lavender-law-armed-conflict/
Methodology source for UN/Gaza MoH revisions, identified records, and problems with women/children proxies. Matched by Priority-A source family: casualty.
Open sourcehttps://www.washingtoninstitute.org/policy-analysis/heres-real-problem-uns-revised-gaza-death-toll
Urban-warfare expert context for Gaza, dense terrain, military difficulty, civilian-risk mitigation, and why simple casualty/destruction metrics are legally weak. Matched by Priority-A source family: loac, intent.
Open sourcehttps://mwi.westpoint.edu/israel-gaza-and-the-looming-challenges-of-urban-warfare/
Retired military assessment of 2021 Gaza conflict, useful for comparing IDF targeting, warnings, and Hamas embedding practices over time. Matched by Priority-A source family: loac.
Open sourcehttps://jinsa.org/wp-content/uploads/2021/10/Gaza-Assessment.v8.pdf
Retired military assessment of prior Gaza operations, useful for Hamas human-shield patterns, IDF precautions, and longitudinal LOAC context. Matched by Priority-A source family: loac.
Open sourcehttps://jinsa.org/wp-content/uploads/2018/10/2014GazaAssessmentReport.pdf
Mainstream methodology source explaining Gaza Health Ministry data limits, identified records, and demographic-reporting changes. Matched by Priority-A source family: casualty.
Open sourcehttps://apnews.com/article/360c6aabc03421c718d4a8452cec2c67
Urban targeting methodology source for weapon choice, tactics, and why blast effects alone do not decide LOAC legality. Matched by Priority-A source family: loac.
Open sourcehttps://lieber.westpoint.edu/targeting-urban-environment-why-weaponeering-tactics-matter/
Counter-audit pairing completed for 3 item(s). Main sources: JINSA (3).
Locator: Counter-audit pairing section of the 2026-06-12 P3 pass.
https://worldagainstisrael.org/method#p3-counter-audit-pairing
UN commission notes reports on ‘Gospel’ and ‘Lavender’, records IDF’s denials, and flags concerns about meaningful human assessment; says it could not independently verify Lavender reporting. Bottom-source disposition: preserves allegation/transmission only; not bottom-source proof. The locator preserves where the allegation appears, not that the allegation is proven. Treat this as source-chain provenance unless separate primary records and counter-audit evidence satisfy the factual and legal threshold.
Locator: A/HRC/56/CRP.4, pp. 39-40, paras. 161-165; Commission summarizes Gospel/Lavender reporting, IDF denial, and its own non-verification caveat.
Quote rule: Verified terms: 'Gospel'; 'Lavender'; 'ISF reportedly denied the claim'; 'The Commission was unable to verify this report independently'.
https://www.ohchr.org/sites/default/files/documents/hrbodies/hrcouncil/sessions-regular/session56/a-hrc-56-crp-4.pdf
Methodology critique of Gaza fatality data, identification status, media-source entries, demographic shifts, and reliability limits. Matched by Priority-A source family: casualty.
Open sourcehttps://www.washingtoninstitute.org/policy-analysis/gaza-fatality-data-has-become-completely-unreliable
Official description: systems generate suggestions; SOPs require human analyst assessment and officer approval; denies autonomous kill decisions.
Open sourcehttps://www.idf.il/en/mini-sites/idf-press-releases-israel-at-war/june-24-pr/the-idfs-use-of-data-technologies-in-intelligence-processing-published-june-18-2024/
Mainstream summary of AP casualty-data findings, useful for public-facing methodology boxes. Matched by Priority-A source family: casualty.
Open sourcehttps://apnews.com/article/e258a4c14641978a00dfb957ce348957
Casualty methodology report on Hamas-run MoH/GMO inconsistencies, combatant/civilian estimates, demographic anomalies, and source-chain risks. Matched by Priority-A source family: casualty.
Open sourcehttps://henryjacksonsociety.org/wp-content/uploads/2025/04/HJS-Hamas-Casualty-Reports-Report-WEB-correct.pdf
Bottom-source audit marker: this dossier includes hostile-source-chain material and must keep source role visible. NGO/UN/media repetition can document how a claim travelled, but the assessment should rest on bottom-source proof, primary records, and counter-audit. Current audit counts: 5 claim-side item(s), 4 allegation-chain-only item(s), 1 primary-like item(s), 17 counter/primary item(s).
Locator: Archive-wide audit rule; generated from docs/codex/BOTTOM_SOURCE_PROOF_AUDIT_2026-06-12.md and JSON queue.
https://worldagainstisrael.org/method#bottom-source-proof-standard
Who first made the concrete allegation?
Did it move through UN, NGO, court, media, or activist channels?
What official, legal, military, or methodology evidence tests it?
Did it become sanctions, lawfare, campus pressure, or media shorthand?
claim_origin
Reported deaths, demographic categories, or civilian-harm totals are used to infer deliberate targeting or criminal intent.
methodology_collapse
The file should separate source custody, named vs aggregate records, combatant uncertainty, demographic distributions, and legal inference.
methodology_audit
Official, UN, NGO, military, and statistical sources should show what the data can support and what it cannot prove.
round15_book_gate
High-confidence misleading: the record supports a serious AI-targeting-methodology question, not the public shortcut that machines autonomously generated civilian kill lists.
round15_book_gate
The claim now has enough public counter/method/legal record to be high-confidence misleading. Guardian/+972/CNN preserve the allegation chain; Lieber Institute, JINSA, Modern War Institute, Washington Institute and IDF-targeting context prevent the public claim from becoming proof of automated unlawful killing.
AI in Gaza: +972/Local Call say ‘Lavender’ mass-marked people with cursory checks; IDF says humans verify and approve. ‘Machines choose civilians’ is an overreach—evidence is disputed; oversight remains contested.