On 28 February 2026, the United States military struck over 1,000 targets in Iran within the first 24 hours of Operation Epic Fury. Between 5,500 and 6,000 targets were hit in total. The system that identified, prioritised, and queued many of those targets was Palantir's Maven Smart System, an AI platform that fused nine separate military intelligence systems into one interface and compressed the kill chain from hours to minutes.
Embedded inside that system was Anthropic's Claude, the same AI model used by developers writing code and students doing homework.
This is not a hypothetical scenario from a policy paper. This is what AI in warfare looks like right now. Algorithms are selecting targets. Machines are compressing the time between identifying a person and killing them. And the companies building these systems are the same ones whose products sit on your laptop.
This article lays out exactly who is deploying AI in warfare, what specific systems they are using, and what the documented consequences have been. Then it asks the question nobody building AI wants to sit with: what does this mean for the rest of us?
Who Is Deploying AI in Warfare Right Now
This is not a future risk. Multiple nations are actively using AI systems in combat operations today. Here are the specific programmes, with documented sources.
Israel: Lavender, Gospel, and Automated Kill Lists
Israel's military has deployed the most publicly documented AI targeting systems in active combat. During the war in Gaza, the Israeli Defence Forces used a system called Lavender to generate targeting recommendations at a scale that would be impossible for human analysts alone.
According to six Israeli intelligence officers who spoke to +972 Magazine, Lavender processed surveillance data, communications intercepts, and social media activity to identify suspected Hamas and Palestinian Islamic Jihad operatives. At one point, the system generated a list of 37,000 Palestinian men flagged as potential targets.
Time for Lavender to identify and approve a target
The numbers that matter:
- During the early weeks of the war, the IDF accepted 15 to 20 civilian deaths for every junior Hamas operative the algorithm identified
- For senior operatives, the accepted civilian casualty ratio was reportedly higher
- Human review of Lavender's recommendations was described by the intelligence officers as a rubber stamp, not a genuine evaluation
Lavender was not the only system. Gospel automatically reviewed surveillance data to recommend bombing targets, primarily buildings and infrastructure. Habsora, introduced in 2021, was built to accelerate target selection using data from drones, satellites, communications intercepts, and social media. The IDF also deployed Alchemist and Red Wolf for surveillance and identification.
The result, as documented by Human Rights Watch, was a dramatic increase in the speed and volume of airstrikes, with a corresponding increase in civilian casualties.
United States: Project Maven and the Minutes-Long Kill Chain
The Pentagon's primary AI warfare programme is Project Maven, now operated by Palantir under a contract ceiling of $1.3 billion through 2029. As of March 2026, Maven Smart System has over 20,000 active users across 35 military tools, with that user base doubling since January 2026.
Maven has reached production-level deployment across nearly every unified combatant command: INDOPACOM, EUCOM, CENTCOM, NORAD/NORTHCOM, SPACECOM, TRANSCOM, AFRICOM, and the Joint Staff.
China: Swarms Without Human Oversight
China's People's Liberation Army is developing AI-enabled drone swarms designed to operate without continuous human control. In one documented test, a Chinese institution supervised approximately 200 autonomous vehicles simultaneously in a coordinated swarm exercise.
China also supplies roughly 80% of the critical technologies used in Russian military drones, according to The Diplomat. Engineers from both nations are collaborating on battlefield adaptation, particularly for loitering munitions, first-person-view strike drones, and fibre-optic guided UAVs designed to resist Ukrainian electronic warfare.
Russia: Scale Through Volume
Russia announced a 200% increase in military spending for 2025-2026, with significant investment in AI-enabled military systems. By mid-2025, Russia was producing approximately 1.5 million FPV drones annually.
Ukraine: Autonomous Drones on the Front Line
Ukraine's Minister of Digital Transformation predicted that 2025 would "significantly increase the percentage of autonomous drones with targeting" and bring "the first real drone swarm uses." Ukraine has been a testing ground for AI-assisted warfare from both sides, with autonomous targeting capabilities evolving rapidly under combat conditions.
The Companies Building This
Here is what has changed in the last two years that most people have not fully registered.
Google removed the restriction on developing weapons or tools for mass surveillance from its code of conduct in February 2025. The same company that in 2018 saw employee walkouts over Project Maven is now providing cloud computing and AI services to the Israeli military under Project Nimbus, a $1.2 billion contract signed in 2021. Amazon is the other primary contractor on Nimbus.
OpenAI signed a $200 million Pentagon deal. This represents a complete reversal from the company's original 2023 usage policy, which explicitly banned military, weapons, and warfare applications. In early 2024, OpenAI quietly removed its ban on military use.
Palantir holds a $10 billion enterprise agreement with the US military. Thirteen former Palantir employees signed an open letter accusing the company of abandoning its founding principles on privacy and human rights.
Anthropic's Claude, the AI model, is embedded inside the Maven Smart System that was used during Operation Epic Fury.
Anduril holds a $20 billion Army contract for AI-enabled military systems.
The Moral Questions We Cannot Avoid
The ethical problems with AI in warfare are not abstract philosophical puzzles. They are specific, documented, and urgent.
The Accountability Gap
International humanitarian law requires that a person be held responsible for civilian deaths in armed conflict. When an AI system identifies a target, recommends a strike, and a human operator approves it in 20 seconds based on information they cannot independently verify, who is responsible for the outcome?
The developer who built the algorithm? The military commander who deployed it? The operator who clicked approve? The company that sold the system?
Currently, the answer is: nobody, clearly. This is not a gap that will be filled by better technology. It is a legal and moral vacuum that grows wider as these systems become more autonomous.
The "Human in the Loop" Fiction
When a system processes 37,000 targets and each one gets 20 seconds of human review, that is not meaningful oversight. That is automation with a human-shaped alibi.
Every military deploying AI weapons claims there is a "human in the loop." The phrase has become a shield against accountability. But the research tells a different story. When AI systems fuse multiple data sources and surface targeting recommendations faster than humans can process, the nominal "human in the loop" becomes an approval node in a process they did not design and may not meaningfully interrogate.
The Misidentification Problem
AI target recognition misidentification rate in complex urban warfare scenarios
In a system processing thousands of targets, a 12.3% error rate translates to hundreds of people incorrectly identified. When the accepted civilian casualty ratio is 15-to-1 on top of that error rate, the mathematics of harm compounds rapidly.
Lowering the Cost of War
This is the implication that should concern everyone, not just those in conflict zones. When AI replaces human soldiers in battlefield roles, it reduces the political cost of waging war. Fewer body bags coming home means less public opposition. Less public opposition means lower barriers to starting conflicts. As one research paper notes, this increases the likelihood of "low intensity" conflicts that risk escalation to broader warfare.
The logic is straightforward: if war becomes cheaper in terms of domestic political consequences, governments will wage more of it. AI does not just change how wars are fought. It changes how often they are started.
How This Affects You
You might think this is a problem for policymakers, military commanders, and people in conflict zones. It is not.
Your Technology Is Dual-Use
The AI models used to recommend bombing targets are built on the same architectures, trained on similar data, and developed by the same companies as the AI you use for writing, coding, and searching the internet. Every subscription, every API call, every cloud hosting bill contributes revenue to companies that are simultaneously building targeting systems. You are not a bystander. You are a customer of the weapons supply chain.
Surveillance Infrastructure Transfers
Human Rights Watch documented that the development, testing, training, and deployment of autonomous weapons systems requires mass surveillance. The surveillance infrastructure built for military AI does not stay in the military. The same facial recognition, pattern-of-life analysis, and predictive targeting techniques have been used to monitor protesters in France, the Occupied Palestinian Territory, and Hong Kong.
The Precedent Problem
Every deployment without accountability sets a precedent. When Israel uses AI to generate 37,000 targets with 20-second human review and the international community does not establish clear legal consequences, it becomes the baseline for the next deployment. The rules being written right now, through action and inaction, will govern how AI is used in every future conflict.
Where International Law Stands
In November 2025, the UN General Assembly's First Committee passed a resolution calling for negotiations on a legally binding agreement on lethal autonomous weapons. 156 nations supported it. Over 120 countries support calls to negotiate a treaty that prohibits or regulates autonomous weapons.
The UN Secretary-General has called for the conclusion of a binding instrument by 2026. Nature published an editorial calling to stop the use of AI in war until laws can be agreed.
But a handful of major military powers, most notably the United States, Russia, Israel, and India, have repeatedly blocked proposals to negotiate binding rules. The same nations deploying these systems are the ones preventing regulation.
What Can Be Done
There are no easy answers. But there are actions that can be taken at multiple levels.
As a citizen: Support organisations working on autonomous weapons regulation. The Stop Killer Robots coalition coordinates international advocacy. Contact your representatives about your country's position on the UN treaty negotiations.
As a developer: Know what your tools are being used for. If you build AI systems, understand that the same capabilities that power useful applications can be repurposed for targeting systems. Consider the military contracting positions of the companies whose platforms and APIs you use.
As a consumer: Recognise that your purchasing decisions fund these companies. This is not a call to boycott all technology. It is a call to be informed about what your money supports.
As a voter: AI in warfare is not a niche issue. It determines how and when your government goes to war, who it surveils, and what accountability exists when things go wrong.
The Question We Are Not Asking
The debate about AI in warfare tends to focus on whether autonomous weapons can be made accurate enough, regulated tightly enough, or controlled adequately enough to be acceptable. That framing misses the deeper question.
The fundamental moral issue is not whether AI can kill precisely. It is whether any system that cannot comprehend the value of a human life should be involved in deciding to end one.
Autonomous weapons systems, as computational systems that are not moral agents, cannot be programmed to assign value to the inherent worth of human life or the significance of its loss. They optimise for objectives. They do not understand what they destroy.
Every system described in this article — Lavender, Maven, Gospel, Lumberjack — is an optimisation engine. It optimises for target identification, kill chain speed, and operational efficiency. What it cannot do is weigh the moral weight of the action it is facilitating.
We, as a society building AI agents that we want to be useful, capable, and reliable, need to confront the fact that the same technology is being used to make killing faster, cheaper, and easier. And until we confront that, the algorithms will keep running, the kill chains will keep compressing, and the body count will keep climbing.
The machines do not have a moral compass. The question is whether we still do.