AI ethics and rules: who answers? · beginner

Are there weapons where AI decides who to kill?

AI in today's wars: what is known about Gaza, Ukraine, Libya and Iran, how solid each report is, who is responsible and where the UN talks on rules stand.

The Palais des Nations in Geneva seen through the Alley of Flags: two rows of national flags on tall poles lead to the building's entrance, marked United Nations, Nations Unies, under a blue sky with white clouds

Photo: Groov3 / Wikimedia Commons

After reading this article you will be able to:

  • Tell apart AI that suggests targets, weapons a person can stop and weapons that attack on their own
  • Judge how solid the reports from Gaza, Ukraine, Libya and Iran are, and what each accused side answers
  • Explain who is responsible under the laws of war and where the UN talks on new rules stand

Drafted with AI from primary sources · checked by a separate AI agent · spot-checked by our editors · How we check

On this page
  1. What has AI done in today’s wars?
  2. Why is this an ethical problem?
  3. Who answers under the law?
  4. Where do the rules stand?
  5. What do the UN, the Red Cross, the Pope and AI companies say?
  6. How to read the next news story about an AI weapon
  7. Where does this answer stop?
  8. Where to go next
  9. Common questions

As of 21 November 2026, no case has been documented in which a weapon chose a person and attacked them with no human decision at all. A few reports come close, from Libya and from Russian drones in Ukraine, but none is confirmed. AI already proposes targets in today’s wars, some drones finish an attack on their own after a person has launched them, and the human check can shrink to a formality. On 25 August 2026 Laurent Gisel of the International Committee of the Red Cross (ICRC), based in Geneva, said there was “no confirmed evidence that fully autonomous weapons had been used to target human beings, but that was the observed trajectory”: things are moving that way (UN summary).

Three expressions in common use, defined in Losing Humanity, a 2012 report by Human Rights Watch and Harvard Law School that called for a ban, help to read every case. The “loop” is the chain of steps from finding a target to striking it.

  • Human in the loop. The machine can pick out a target but attacks only when a person orders it: the AI proposes, a person approves.
  • Human on the loop. The machine selects and attacks by itself while a person watches and can stop it.
  • Human out of the loop. Once started, the machine selects and attacks with no person stepping in.

For the ICRC, autonomous weapons “select and apply force to targets without human intervention” (position). In the weapons cases below (Ukraine, Russia, Libya), a person still launches the weapon and sets an area, a route or a kind of target; what can be missing is a human decision on the exact target and moment. In Gaza and Iran, software proposes targets and people carry out the strikes.

Each case is labelled confirmed (an official document, or two independent investigations with different sources), reported (one investigation or anonymous sources) or disputed (sources contradict each other). The accused side’s answer is given whenever one was found, death tolls name who counted them, and the page does not judge who is right in any war.

This site is written with AI tools, including Claude by Anthropic, and Claude is part of this story: two newspapers reported that it was used with US military targeting software in the 2026 strikes on Iran, and Anthropic is in a legal dispute with the Pentagon (the US Department of Defense) over military use of its AI. This page holds Anthropic, Google and OpenAI to the same standard and uses only published sources, linked in the text.

What has AI done in today’s wars?

CaseWhat the machine did, according to the sourcesWho decidedHow solid
Gaza, from 2023Suggests targets to analystsPeople approve, says the army; how carefully is disputedUse confirmed; details disputed
Ukraine, from 2024Steers drones in the last stretchA person, says the governmentConfirmed as official position; exceptions reported
Zaporizhzhia, July 2026A Russian drone was to choose its final target; it crashed, killing three people (NYT)No one in the last stage, say Ukrainian investigatorsReported
Libya, 2020Drones a UN panel called “autonomous”Unclear; the maker says a personDisputed
Iran, 2026Software proposed targetsPeople, says the PentagonUse confirmed; role in a school strike disputed

Gaza

The cases here come from the war in Gaza that followed the attack on Israel by Hamas and other Palestinian armed groups on 7 October 2023 (UN News).

Confirmed: the Israeli army uses AI tools that suggest targets. On 2 November 2023 it said its intelligence directorate was using artificial intelligence and automated tools to “produce reliable targets quickly and accurately” (Times of Israel). In June 2024 it said one system, Habsora (“The Gospel”), has been in use since 2021 and “suggests to the intelligence analyst” what to research (archived IDF statement).

Reported: in April 2024 six anonymous Israeli intelligence officers spoke to +972 Magazine and Local Call, an Israeli-Palestinian magazine and its Hebrew-language partner, and through them to The Guardian. Four of them said a system called Lavender had listed “as many as 37,000” Palestinians as suspected militants; +972 reports errors “in approximately 10 percent of cases”. The sources said a tool called “Where’s Daddy?” alerted officers when a marked person was at home, so that the home could be struck. One officer said he spent about 20 seconds on each target. Human Rights Watch confirmed that the army uses a tool that tracks people in this way, but could not confirm other technical details.

The army’s answer: the Israel Defense Forces (IDF) told The Guardian it “does not use an artificial intelligence system that identifies terrorist operatives or tries to predict whether a person is a terrorist.” It called the system a database, “not a list of confirmed military operatives eligible to attack.” Its June 2024 statement calls Lavender “a general-purpose database” and says attacks are approved by commanders, “not intelligence analysts, and certainly not technological tools.” It told +972 that “each target is examined individually”, and told AP in 2025 that analysts examine the targets AI helps identify independently, with senior officers.

Disputed: what Lavender does, and how real the human check was. No source says a machine fired on its own in Gaza.

Ukraine and Russia

Confirmed, as an official position: AI steers drones, mostly in the last stretch. In Ukrainian systems described by Reuters, a pilot selects a target on the camera and software completes the approach. Ukraine’s Ministry of Defence said on 18 August 2026 that more than 70 systems using AI and computer vision, software that recognizes things in camera images, do tasks up to “autonomous target engagement” (striking with no person steering), but “the final decision on target selection and engagement is made by a human operator.”

Reported: exceptions. Mykhailo Fedorov, who was defence minister until shortly before, told The New York Times in late July 2026 that Ukraine was testing in Crimea AI systems that pick targets themselves; Meduza, summarizing the Times, adds that no deaths from those attacks have been reported (Meduza). On the Russian side, Ukraine’s military intelligence says the V2U drone can “autonomously search for and select targets”, and a report by CSIS, a Washington research centre, based on Ukrainian analysis, concludes that Russia “has likely fielded a fully autonomous unmanned system in combat”.

Reported: the Zaporizhzhia drone. According to The New York Times of 24 August 2026 (republished), a Russian Molniya drone hit Zaporizhzhia on 6 July. It carried an Nvidia Jetson Orin, a small computer that runs AI on board. Its operators had programmed it to fly toward a gas station, where it was to choose its exact target by itself, most likely propane tanks, experts and military officials told the Times. “As it tried to navigate toward a gas station, it crashed into a wall and exploded”, killing a 19-year-old student and two other people. Nothing in the Times’ account says the drone picked out those three people.

The evidence comes from Ukrainian investigators: the drones had no antennas to talk to an operator and sent no radio signals, and their code showed what they had been trained to attack. Nvidia says it does not sell the module in Russia, adding that it is widely available through resellers. No Russian official comment on these claims was found as of 21 November 2026.

Libya, 2020

Disputed. A UN panel of experts on Libya says retreating forces were “hunted down and remotely engaged” by drones or “lethal autonomous weapons systems such as the STM Kargu-2”, “programmed to attack targets without requiring data connectivity between the operator and the munition”, that is, without a radio link to the person controlling them. The report cites a confidential source, does not say whether a given attack was made by the drones or by the autonomous weapons, and does not say that anyone was killed by them. Turkey’s defence industry chief Ismail Demir denied it: “Navigation to the target is also done under human control.” The maker’s product page says the strike “is fully performed by the operator”.

The United States and Iran, 2026

Confirmed: in May 2026 the Pentagon’s chief digital and AI officer said that Operation Epic Fury, the US air campaign against Iran, used Palantir’s Maven Smart System, software that brings together military data and helps choose targets, “to conduct strike missions across the entire battle space, 13,000 targets in 38 days”, and that “you always have the human who will analyze the situation”.

Reported by two newspapers: Claude was used with Maven. The Washington Post reported on 4 March 2026, citing two people, that “Maven, powered by Claude, suggested hundreds of targets” and ranked them (republished by Stars and Stripes); Anthropic and Palantir declined to comment. On 11 March The New York Times wrote that Claude “does not directly create targets” but works with Maven and other software to “identify points of interest” (republished). Disputed: how directly Claude shaped the choice of targets. In June 2026 Anthropic’s chief executive, Dario Amodei, asked by Bloomberg about the Minab strike, said “we don’t know exactly how these models were used” and that this use did not break the company’s rules (Bloomberg).

Disputed: the school in Minab. On 28 February 2026 an elementary school in Minab, Iran, was hit. Iranian officials put the dead at “at least 175 people, most of them children”, the Times reported; a UN fact-finding mission later cited “credible independent information” of 157 dead, 123 of them children (Stars and Stripes). A preliminary US military inquiry, reported by the Times in March, found the United States responsible and blamed outdated intelligence data and, officials said, likely human error rather than new technology.

In September, Bloomberg cited officials involved in an unpublished Pentagon review: the school, still listed in US databases as a military site, came out of Maven as a recommended target for the first day of strikes, and some staff relied on the software too much. That was one of three failures they described, with bad intelligence and outdated satellite images; Bloomberg’s reporting also points to time pressure and to cuts in the teams that check harm to civilians (via Gizmodo). Palantir says it “is not responsible for the underlying data nor identifying intelligence deficiencies” and that there is no evidence its software was at fault. A senior administration official told Bloomberg: “The United States does not target civilians.” As of 21 November 2026 the review was still unpublished, and CENTCOM, the US command for the Middle East, had not answered questions on its status (Stars and Stripes, 19 November).

On 17 September the UN mission said it had “reasonable grounds to believe” US forces were responsible and that the strike was an indiscriminate attack and a war crime, pointing to the failure to update the school’s information in US targeting databases (Reuters, via Internazionale). The same report found that the Iranian authorities committed crimes against humanity in their crackdown on anti-government protests from late December 2025, a finding unrelated to AI; Iran has consistently rejected allegations of systematic rights violations, Reuters notes. The mission reached its conclusions without answers from the US government, Stars and Stripes reports. The Pentagon declined to comment on the report and said “Our investigation remains ongoing”; the White House said Iran was the only party in the conflict that had committed war crimes. In April, before the report, a group of former senior US military legal advisers had written that it was “premature and speculative” to call the strike a violation of the laws of war or a war crime, Stars and Stripes notes. No version says the strike was launched without a human decision.

Why is this an ethical problem?

Human dignity. Christof Heyns, an independent expert appointed by the UN, wrote in 2013 that “machines lack morality and mortality, and should as a result not have life and death powers over humans” (report).

Trusting the machine too much. The ICRC warns of automation bias, “a propensity to rely on machine outputs even when other available information may call those outputs into question”, which can turn a person into “a human rubber stamp” who approves without really checking (2025 submission). In Iraq in 2003, US Patriot air defences shot down two coalition aircraft, killing three crew; a Pentagon advisory board found, among other failures such as a poor friend-or-foe identification system, that operators were “trained to trust the system’s software”.

Experiments with the general public show the tendency. In a peer-reviewed 2024 study, people in a simulated drone-strike memory game reversed their choice on whether to strike in 62% of cases (135 students) and 67% (423 US adults online) when a robot that in fact answered at random disagreed. The authors say the task “bears no resemblance to real-world military threat-detection procedures”, and the decision had no time limit. In a peer-reviewed study of about 9,000 adults in nine countries who had 7 or 10 seconds to identify aircraft, about one answer in four changed after a suggestion that came, at random, from an AI or a human analyst and was, at random, right or wrong.

The first studies with military participants point the other way. A peer-reviewed study published in July 2026, read here only in its summary, gave 2,015 Israeli military personnel a replica of an AI targeting support system and found “strong evidence of algorithmic aversion”, that is, distrust of the machine, especially when much harm to civilians was expected. In a preprint, not yet peer-reviewed, 236 West Point cadets trusted AI advice in a more balanced way than the public. None of these studies observed real operations.

Speed, and wars that become easier to start. A 2025 report by SIPRI, a Stockholm institute that studies weapons and conflicts, calls speed “a key strategic driver” and “a key source of risk”, with operators at risk of becoming “passive approvers”. Heyns warned that such weapons may “lower the threshold for States for going to war or otherwise using lethal force”. Political scientist Michael Horowitz, who until 2024 worked on US autonomous weapons policy at the Pentagon, notes that “indirect artillery and manned airpower” (guns firing at targets out of sight, and planes with pilots) once raised similar fears. In that 2016 essay he judges autonomous weapons unlikely to raise new problems about the decision to go to war, apart from the risk that they make it too easy.

Who answers under the law?

People and states, never the machine. The laws of war, in the Geneva Conventions and their Additional Protocol I of 1977, a treaty added to them, say that:

  1. attacks must distinguish civilians from fighters (Article 48);
  2. an attack is banned if the harm expected to civilians is excessive compared with the military advantage expected (Article 51);
  3. those who “plan or decide upon an attack” must “do everything feasible to verify” that the target is a military objective (Article 57);
  4. states that have joined the Protocol must check every new weapon against the law (Article 36).

In 2019 the states of the Convention on Certain Conventional Weapons (CCW), an arms treaty that already limits weapons such as landmines and blinding lasers, agreed: “Human responsibility for decisions on the use of weapons systems must be retained since accountability cannot be transferred to machines” (guiding principles). Their experts’ 2026 report repeats it: “IHL imposes obligations on States, parties to armed conflict, and individuals, not on machines” (IHL is international humanitarian law, the laws of war).

The disagreement is about harm that no one wanted and no one could foresee. Heyns warned that “a responsibility vacuum will emerge”. Human Rights Watch, which campaigns for a ban, wrote in 2012 that there is “no fair and effective way to assign legal responsibility”. In 2013 two US Naval War College lawyers called that “a false premise”: people decide how to program the weapon and when to launch it, and a commander answers if he or she knew or should have known that it was programmed to commit war crimes and did nothing. Human Rights Watch replied in 2015 that commanders would still likely escape justice in the more common case in which they could not foresee or stop an unlawful attack.

Where do the rules stand?

In Geneva. The CCW’s group of government experts adopted its final report by consensus on 4 September 2026, with over 100 states taking part; consensus means no state present objected. It is not a treaty but “a set of elements of an instrument, without prejudging its nature”: a list of points for future rules, without deciding yet whether they will be binding. It describes lethal autonomous weapons systems (LAWS in UN documents) as able to identify, select and engage a target “without intervention by a human being operating the system”, and says “control and human judgement with regard to LAWS are needed”. Campaigners and some states ask for “meaningful human control”, a term coined in 2013 by the British group Article 36 (UNIDIR). The final text does not explicitly ban weapons designed to attack people; a ban on weapons whose effects “cannot be anticipated and limited”, in the draft of 3 September, was dropped.

In New York. The UN General Assembly has voted every year since 2023:

VoteYesNoAbstainedVoted no
Resolution, Dec 2023152411Belarus, India, Mali, Russia
Resolution, Dec 2024166315Belarus, North Korea, Russia
Resolution, Dec 202516467Belarus, Burundi, Israel, North Korea, Russia, United States
Draft, disarmament committee, 29 Oct 202616359Belarus, Israel, North Korea, Russia, United States

The 2026 draft calls on CCW states “to advance to negotiations on an instrument”, a set of rules, binding or not. The Secretary-General had asked for a binding treaty by 2026 banning weapons that work without human control or oversight and cannot respect the laws of war (report). As of 21 November 2026 there is none, and in Geneva states reportedly chose to “elaborate” rules rather than “negotiate” them (see below).

United States. Directive 3000.09, the Pentagon’s rule on autonomy in weapons, last updated in January 2023, requires weapons that let commanders and operators “exercise appropriate levels of human judgment over the use of force”; the Congressional Research Service notes that “U.S. policy does not prohibit the development or employment of LAWS”. A presidential memorandum of 5 June 2026 ordered an update within 90 days, to be reviewed every year. The deadline has passed, and as of 21 November 2026 no new version had been made public (DroneXL, 16 November).

European Union. The AI Act does not apply to AI used “exclusively for military, defence or national security purposes” (Article 2(3)).

Italy. In 2024 Italy asked for a new CCW protocol banning weapons “operating completely outside human control and a responsible chain of command”. For a weapon to be regulated rather than banned, it says a person must keep at least “the final crucial function, which is the decision of whether or not to apply force to a previously identified and selected target”. In September 2026 it welcomed the consensus report without saying which next step it prefers. At the November conference its delegation called “for a strong mandate for a new protocol”; no Italian comment on the outcome was found as of 21 November 2026.

What did the Geneva conference decide?

The CCW’s Seventh Review Conference, held every five years to decide the treaty’s future work, ended in the early hours of 21 November 2026. The proposals on the table went two ways: experts would “negotiate a new Protocol”, an addition to the treaty (proposal), or, as the United States proposed, write “version 1.0 of the non-binding instrument by 2029”, rules that countries would follow voluntarily. Decisions need consensus, so one state can block them.

According to the news agencies AFP and AP, states agreed to “elaborate” an instrument on these weapons, that is, to work out a set of rules, rather than to “negotiate” one. The conference president, Swiss ambassador Julien Thöni, said afterwards, as reported by AFP, that some countries had objected to the word “negotiate” and that the work should be done “over the next two years”; AP gives 2028 as the target. Whether the rules will be binding is not decided: AP writes that states “didn’t commit to binding rules”. As of 21 November 2026 the official text had not been published: the last official draft, of 20 November, leaves the decision on autonomous weapons blank.

What do the UN, the Red Cross, the Pope and AI companies say?

The UN Secretary-General calls machines that can take lives without human control “politically unacceptable, morally repugnant” (May 2025). With the ICRC President he warned in August 2026 that the world is “dangerously close to crossing a moral red line: the autonomous targeting of humans by machines.”

The ICRC wants binding rules that ban unpredictable autonomous weapons and their use against people, and set limits on all the others (position); before the conference it asked states to “launch negotiations on a new Protocol” (June 2026). No ICRC comment on the outcome was found as of 21 November 2026.

Pope Leo XIV wrote in his encyclical, a letter to the whole Church, Magnifica Humanitas (May 2026, paragraph 198) that “it is not permissible to entrust lethal or otherwise irreversible decisions to artificial systems”. On 30 October 2026, before the conference, he encouraged the states about to meet in Geneva “to take concrete steps toward banning lethal autonomous weapons”.

Campaigners and states. The Stop Killer Robots campaign calls a binding treaty “the only acceptable next step”; after the conference it said, as quoted by AFP, that “States adopted a watered-down mandate that fails to chart a clear path towards a legally binding instrument.” The United States says autonomy used well can “reduce the risks to civilians”. In views sent to the UN in 2024 (report), Israel says these weapons “may serve both military necessity and humanitarian considerations”, and Russia sees “no convincing grounds for imposing any new limitations”.

AI companies, rules in force on 21 November 2026:

  • Anthropic’s usage policy, in force since 12 November 2026, bans using its AI to develop or operate software for weapons, “including targeting, fire control, or engagement”, and to put weapons on drones or other “unmanned or autonomous platforms”; it lists “battlefield management” among banned surveillance uses. It allows contracts with some government customers that “tailor use restrictions” to their mission, if Anthropic judges the safeguards adequate. The previous version, in force since 15 September 2025, already banned using its AI to design weapons.
  • Google removed in February 2025 its list of uses it would not pursue, which included weapons (PA Media/dpa, via The Star); its principles now promise “appropriate human oversight”. In April 2026 it agreed to let the Pentagon use its Gemini AI on classified networks; a US official told NBC News the deal covers lawful use, and Google said AI “should not be used for domestic mass surveillance or autonomous weaponry without appropriate human oversight”.
  • OpenAI’s usage policies, in force since 20 November 2026, ban “weapons development, procurement, or use” and national security uses “without our review and approval”, which it may grant to specific partners case by case, always keeping “human judgment and accountability over high stakes decisions, including the use of force”. Its principles for national security partnerships say targeting-related uses “are not categorically disallowed” but need “heightened scrutiny”, and that it does not support systems that “autonomously identify, select, and engage targets without appropriate levels of human authorization”. In January 2024 it had dropped its ban on “military and warfare” uses (The Intercept). Its 2026 agreement with the Department of War, a name the US government has also used for the Pentagon since September 2025 (executive order), allows use “for all lawful purposes”, but says its AI “will not be used to independently direct autonomous weapons” where “law, regulation, or Department policy requires human control”. In February 2025 an AP investigation found that the Israeli army’s use of AI from Microsoft and OpenAI, whose models it bought through Microsoft’s cloud, rose sharply after 7 October 2023; OpenAI said it has no partnership with Israel’s military, and Microsoft declined to comment.

Anthropic and the Pentagon. In February 2026 Anthropic refused two uses in its Pentagon contract: mass domestic surveillance and fully autonomous weapons, those that “take humans out of the loop entirely”. It said today’s AI is “not reliable enough” for fully autonomous weapons, while calling partially autonomous weapons “vital to the defense of democracy” and adding that fully autonomous ones “may prove critical for our national defense”. The Pentagon’s top spokesman wrote on social media, AP reported, that the military has no interest in mass surveillance of Americans or in autonomous weapons “that operate without human involvement”, but wanted to use Anthropic’s AI “for all lawful purposes”. The Pentagon labelled Anthropic a supply chain risk, a status usually given to firms suspected of working for hostile governments, which lets it bar staff and contractors from using the company’s products for defence work. The Pentagon relied on two laws. In August 2026 a federal court in California ruled for Anthropic and cancelled the designation under one of them; on 25 September the appeals court in Washington upheld it under the other, by two votes to one (Nextgov; Congressional Research Service). In October Anthropic asked the full appeals court to hear the case again (MLex); on 18 November the government asked it to refuse, Law360 reported. An appeal in the California case was opened in late October (court record). No decision on the rehearing was found as of 21 November 2026.

How to read the next news story about an AI weapon

Four questions help:

  1. What did the machine do: suggest a target, steer the last stretch, or choose and attack by itself?
  2. Who says so: an official document, two independent investigations, or one source without a name?
  3. Has the accused side answered, and what did it say?
  4. Who counted the dead, and does anyone else confirm the number?

Where does this answer stop?

  • No independent access. No independent investigator has examined how Lavender, Maven or the Russian drones work; much of what is known comes from anonymous sources or one side of a war.
  • Articles for subscribers only. The New York Times, Washington Post and Bloomberg articles were read in authorized republications, free opening lines or other outlets’ reports, named in each link.
  • Documents not seen. The Pentagon’s Minab review is not public; the full UN mission report on Iran could not be opened; the 2026 study with Israeli military personnel was read only in its summary.
  • The Geneva decision is known only from AFP and AP reports: the official text had not been published on 21 November 2026, and its exact wording may differ.
  • No verdicts. This page does not decide whether any strike was lawful.

Where to go next

Common questions

Do "killer robots" exist?

Not in the sense of a machine that decides by itself to find and attack people: as of 21 November 2026 no such case is documented. Weapons that attack objects on their own within limits set by people existed before today's AI, such as air defences against missiles and the Israeli Harpy drone, which seeks out radars and, according to a 2017 report by SIPRI, a Stockholm institute that studies weapons and conflicts, "operates in complete autonomy". What is new is AI that proposes targets, and reports from Ukraine of drones that choose the final target themselves. "Killer robots" is the name used by the Stop Killer Robots campaign.

Are autonomous weapons illegal?

No treaty bans them as such. The laws of war apply to them fully: attacks must not be aimed at civilians or cause excessive harm to them, and states that have joined Additional Protocol I, a 1977 treaty added to the Geneva Conventions, must check every new weapon against the law. In September 2026 the states taking part in talks under a UN arms treaty agreed by consensus on the elements of possible new rules. At a conference in Geneva that ended early on 21 November 2026, they agreed, according to the AFP and AP news agencies, to "elaborate" rules on these weapons over about two years, without deciding whether they will be binding. The official text had not been published that day.

What is Lavender, the AI system used in Gaza?

According to six anonymous Israeli intelligence officers quoted by +972 Magazine and Local Call in April 2024, Lavender is an AI system that, according to four of them, marked as many as 37,000 Palestinians as suspected militants; +972 reports errors in about 10% of cases. The Israeli army says Lavender is a database, "not a list of confirmed military operatives eligible to attack", and that attacks are approved by commanders, not by software. The two versions have not been checked independently.

Did AI choose the school that was hit in Minab, Iran?

It is disputed. The school was hit on 28 February 2026; Iranian officials said at least 175 people died, most of them children, and a UN fact-finding mission later cited information of 157 dead, 123 of them children. A preliminary US military inquiry reported by The New York Times in March found the United States responsible and blamed outdated intelligence data and, officials said, likely human error. In September Bloomberg, citing officials involved in an unpublished Pentagon review, reported that the school came out of Palantir's Maven software as a recommended target and that some staff relied on it too much, alongside bad intelligence and old satellite images; Palantir says it is not responsible for the data. No version says the strike was launched without a human decision.

Do Anthropic, Google and OpenAI let the military use their AI?

Yes, with different limits. Anthropic bars its AI from weapons software, including targeting, but allows tailored contracts with some governments; two newspapers reported that Claude was used with Maven targeting software in US strikes on Iran in 2026; it refused fully autonomous weapons and mass domestic surveillance, and the Pentagon labelled it a supply chain risk, a case still in court. Google dropped its pledge not to build AI for weapons in February 2025 and in April 2026 let the Pentagon use its AI on classified networks, saying autonomous weapons need appropriate human oversight. OpenAI bans weapons uses and allows national security uses only after its review, case by case; AP found in 2025 that Israel's army used OpenAI models bought via Microsoft's cloud; its 2026 Pentagon deal bars its AI from independently directing autonomous weapons where US rules require human control.

Topics: Who answers?, autonomous weapons, AI ethics, international law, United Nations

Diagram