By Stuart Kerr, Technology Correspondent, LiveAIWire

AI Gaza damage assessment tools have produced one of the starkest numbers to come out of any conflict this decade: as of October 2025, satellite analysis found that 81 percent of all structures in the Gaza Strip showed some level of damage, with 123,464 buildings classified as fully destroyed. That figure comes from UNOSAT, the UN’s satellite imagery programme, which has published a running series of comprehensive damage assessments throughout the war using deep learning applied to commercial satellite imagery. It is one of two very different ways artificial intelligence has entered this conflict, and the two uses raise almost opposite questions.

The first use is documentation: AI helping outside observers measure, at scale and independent of any single party’s account, what has physically happened to a city. The second use is targeting: AI-assisted systems reportedly helping Israeli military intelligence identify and prioritise bombing targets during the war. Both are genuinely contested in different ways, and treating them as one story risks getting both wrong.

What AI Gaza Damage Assessment Data Actually Shows, and What It Cannot Show

UNOSAT’s methodology compares satellite images taken at different points in the conflict against pre-war baseline imagery, using automated and semi-automated classification to flag structures as destroyed, severely damaged, moderately damaged or possibly damaged.

The October 2025 assessment, UNOSAT’s own documentation states plainly, is “a preliminary analysis and has not yet been validated in the field,” a caveat the agency has attached to every update in the series. A separate joint Rapid Damage and Needs Assessment published by the World Bank, the European Union and the United Nations in April 2026 used artificial intelligence to examine the length of Gaza’s road network from satellite imagery, categorising roughly 74 percent of the network as destroyed, while explicitly noting that “due to limitations from the visual assessment method, these figures cannot be fully verified.”

That joint assessment put total physical damage to Gaza’s infrastructure at approximately 35.2 billion dollars as of the October 2025 ceasefire date, with housing accounting for the largest single share at 18 billion dollars, and estimated recovery and reconstruction needs at 71.4 billion dollars. These are large, specific, cross-checked figures, produced by three independent institutions working jointly rather than any single government or advocacy group.

What they measure is physical and economic damage. What they do not and cannot measure, and what none of the researchers behind them claim to measure, is legal responsibility, military necessity, or intent behind any individual strike. Establishing those things is the job of courts and investigators working from a different evidentiary standard entirely, not satellite pixels.

The Targeting Systems: What Is Documented, What Is Disputed

The second AI story in this conflict concerns systems reportedly used inside Israeli military intelligence to help select targets, most prominently ones that have been reported under the names Lavender and Habsora. According to a 2024 investigation by +972 Magazine, an Israeli-Palestinian outlet, based on interviews with six unnamed Israeli intelligence sources, Lavender assigned Palestinians a probabilistic score indicating the likelihood they were militants, and at one point flagged approximately 37,000 people as suspected targets, with sources describing an acknowledged error rate of around 10 percent in that process.

The same reporting described Habsora as a separate system used to identify buildings and infrastructure as targets, generating recommendations at a pace one former intelligence officer characterised to reporters as far faster than manual target selection.

The Israeli military disputes significant elements of this account. In a statement to the Guardian, cited in reporting on the story, the military said it “does not use an artificial intelligence system that identifies terrorist operatives” and described Lavender instead as “simply a database whose purpose is to cross-reference intelligence sources.” The military’s public position is that human analysts independently verify any AI-assisted recommendation against the legal requirements of international law before a strike is authorised, and that final targeting decisions are made by human officers, not by software.

Independent verification of exactly how much scrutiny individual recommendations received in practice, and how consistently that scrutiny was applied across tens of thousands of cases, is not something outside researchers have been able to establish conclusively, precisely because the relevant internal records are not public.

Why the Two AI Stories Point in Different Directions on Accountability

The documentation story and the targeting story sit in tension with each other in an interesting way. Satellite-based damage assessment exists specifically to reduce the space for any party to a conflict, including outside advocacy groups, to assert unverified figures without independent evidence. It functions as a check on claims, not a source of them.

Algorithmic targeting assistance, by contrast, has generated exactly the kind of accountability concern that independent verification is supposed to resolve: a process where, according to the disputed sourcing above, human oversight may in practice have been thinner than the formal chain of review implies, and where the specific line between an AI system recommending a target and a human meaningfully evaluating that recommendation is difficult for anyone outside the process to audit.

That tension is not unique to this conflict. LiveAIWire’s earlier coverage of the ethical questions raised by AI in military strategy generally found that the International Committee of the Red Cross has argued fully autonomous weapons systems cannot reliably make the context-sensitive legal judgements international humanitarian law requires, a position several governments developing these systems dispute on the grounds that specific systems can in principle be validated against those standards.

Whether Lavender and Habsora constitute the kind of automated decision-making the ICRC’s position addresses, or whether they are simply data-processing tools inside a fully human-controlled process as the Israeli military describes them, is precisely the question that remains unresolved, both in this conflict and in the broader international debate over AI-assisted targeting. LiveAIWire’s related reporting on how AI is reshaping combat more broadly found the same accountability gap recurring across autonomous weapons debates generally: when a targeting recommendation comes from software, responsibility for the outcome does not transfer to the machine, whatever the formal division of labour between analyst and algorithm turns out to have been in any specific case.

The Legal Proceedings Running in Parallel

None of this technical debate exists separately from an active legal process. The International Criminal Court issued arrest warrants in November 2024 for Israeli Prime Minister Benjamin Netanyahu and former defence minister Yoav Gallant, with the Court stating there were reasonable grounds to believe both men bore criminal responsibility for the war crime of using starvation as a method of warfare and for crimes against humanity including murder and persecution, in connection with Israel’s military campaign in Gaza.

That warrant is a judicial decision that a case may proceed to trial. It is not a criminal conviction, and Netanyahu denies the allegations. Israel disputes the ICC’s jurisdiction over the matter, noting that Israel, like the United States, is not a party to the Rome Statute that established the court, though the ICC’s jurisdiction in this case rests on the location of the alleged conduct rather than Israel’s own membership.

The warrant remains a live diplomatic issue as of July 2026. New York Mayor Zohran Mamdani stated this month that his administration had concluded the city lacks independent legal authority to execute the ICC warrant should Netanyahu travel to New York for the UN General Assembly in September, while calling on the US federal government to do so instead. President Trump responded on social media that Netanyahu “will not be arrested, in any way, shape, or form, while in the United States.” Israel’s ambassador to the UN, Danny Danon, criticised Mamdani’s statement directly.

More than 120 countries that are parties to the Rome Statute carry a formal obligation to arrest Netanyahu should he enter their territory, though enforcement in practice has varied, and neither the United States nor Israel is among those states parties.

What This Means for How AI-Assisted Conflict Reporting Should Be Read

The practical lesson from laying both AI stories side by side is that neither supports a simple verdict on its own, and conflating them tends to produce exactly the kind of overreach that undermines otherwise solid evidence. The UNOSAT and RDNA damage figures are independently verified, cross-institutional, and explicit about their own limitations, which is what makes them credible as documentation. The claims about Lavender and Habsora rest substantially on anonymous sourcing that the subject of the reporting directly disputes, which does not make those claims false, but does mean they carry a different evidentiary weight than a satellite-derived structural count, and should be read and cited accordingly.

The ICC warrant sits in a third category entirely: a formal legal finding, made under a defined evidentiary standard by an international judicial body, that a case should proceed, which is neither proof of guilt nor something that should be dismissed as a mere accusation with no institutional weight behind it. Readers trying to understand accountability in this conflict are better served by holding these three categories of evidence, independently verified physical documentation, contested intelligence-sourced reporting, and an active judicial process, separately, rather than treating any single one as settling what the others have not yet resolved.

Why AI Gaza Damage Assessment Methods Are Likely to Become Standard in Future Conflicts

The methodology UNOSAT and the joint World Bank, EU and UN team applied in Gaza did not originate there. Satellite-based damage classification using machine learning has been used, in less comprehensive form, to document destruction in Ukraine, Syria and Sudan, and the scale and granularity of the Gaza assessments are likely to set a working standard other conflict-monitoring efforts will be measured against.

That has a genuine upside for anyone trying to understand a conflict from outside it: a technique that does not depend on access to a war zone, and does not depend on any single government’s cooperation, can keep producing data even when journalists, investigators and aid workers cannot safely enter an area.

It also has a real limitation worth stating plainly rather than glossing over. AI Gaza damage assessment work answers the question of what was destroyed. It does not and cannot answer the question of why, under what rules of engagement, or with what degree of care to avoid it, any specific building came down. Those questions require testimony, forensic reconstruction and legal process, the kind of evidence an international court is built to weigh, not the kind an image classifier produces.

Treating satellite-derived destruction figures as a substitute for that legal process, in either direction, either as proof of a crime or as something that can be waved away because it does not itself establish intent, misunderstands what the technology is actually for.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.