AI in War: What the Iran War Reveals About the Pentagon’s Algorithms

On Feb. 28, 2026, the United States and Israel launched a military campaign against Iran, striking more than 900 targets in the first 12 hours and killing Iran’s Supreme Leader Ayatollah Ali Khamenei. The conflict is still raging, with strikes continuing across the country and the region destabilising by the day. Yet behind the missiles and fighter jets lies another revolution in how this war is being fought.

AI, the same technology that millions use daily to draft emails or summarise documents, has become a central instrument of lethal military power. Anthropic’s Claude AI model is embedded inside the Pentagon’s targeting and intelligence apparatus, processing satellite imagery, intercepted communications, and operational data to help commanders decide who to strike, where, and when.

What once required days of human analysis is now compressed into hours or minutes, enabling a pace of warfare that no prior generation of military planners could have executed. AI has been present on battlefields before, from drone guidance systems to satellite image analysis, but the Iran conflict represents its most expansive and consequential deployment to date, and the full implications of that scale are still unfolding.

To understand what is happening in Operation Epic Fury, it is necessary to understand what AI is actually doing inside the U.S. military’s targeting process. The Pentagon had been building toward this moment since 2018, when it launched Project Maven, a program designed to integrate machine learning into military intelligence and targeting. By 2026 that program had evolved dramatically.

Through a partnership with the data-analytics company Palantir, Claude serves as the reasoning engine inside a decision-support system embedded across all US combatant commands. The system allows planners to organise kill chains and orchestrate the movement and actions of hundreds of military units in hours rather than days. During the 2003 invasion of Iraq, a 2,000-strong intelligence unit was needed to handle target identification. In Operation Epic Fury, 20 troops are carrying out the same workload.

This compression of planning time is what military analysts call decision compression. AI simultaneously ingests drone footage, signals intelligence, logistics data, and social media streams to surface ranked target lists alongside weapon recommendations and automated legal assessments. The result so far has been more than 1,250 targets struck in the first three days alone, including ballistic missile sites, naval vessels, and air defence systems.

Alongside AI-driven targeting, the operation has introduced a new weapons platform to the battlefield. The Pentagon deployed Low-Cost Unmanned Combat Attack System drones, known as LUCAS, for the first time in combat. At $35,000 per unit, each LUCAS costs roughly a fiftieth of a single Tomahawk cruise missile. These loitering munitions, modelled on Iran’s own Shahed drones, can hover over a target area before diving onto their mark with precision.

Their deployment reflects the doctrine of affordable mass, which is the strategy of saturating enemy defences with high volumes of cheap, AI-guided weapons rather than relying exclusively on expensive precision munitions. Cyber operations are running in parallel with the strikes, including the hacking of a popular Iranian religious app used by over 5 million people and the repeated disabling of internet connectivity across the country. Together, the drone swarms, the cyber operations, and AI-guided targeting represent a new template for American warfare.

Nevertheless, the speed and scale that make AI militarily attractive also carry serious risks the conflict is already exposing. On Feb. 28, a missile struck near a school in southern Iran, killing 165 people, many of them children. The site appeared adjacent to a military barracks, and the United Nations described the strike as “a grave violation of humanitarian law.” It remains unknown what role, if any, AI played in selecting or approving that target.

That uncertainty is itself the defining problem. When a human commander has only minutes to review an AI-generated strike recommendation backed by synthesised data they could not have assembled themselves, the decision is less an independent judgment and more an approval. AI systems are also known to be fallible in unpredictable environments, and targeting systems used in prior conflicts have demonstrated error rates of approximately ten percent, raising the prospect that in a campaign of this scale, hundreds of targets may have been flagged in error.

The military operation unfolded against the backdrop of a political confrontation that exposed how fragile the frameworks governing AI in war truly are. Until the final week of February 2026, Anthropic held an exclusive Pentagon contract worth up to $200 million to deploy Claude across classified military networks. That contract, originally negotiated under the Biden administration, included two explicit restrictions: Claude could not be used for mass surveillance of American citizens, and it could not be used to control fully autonomous weapons without meaningful human oversight. The Trump administration had initially accepted those terms. The rupture began when Anthropic started asking detailed questions about how its technology had been used in the January 2026 operation that led to the capture of Venezuelan President Nicolas Maduro, and from there, the confrontation escalated rapidly.

U.S. Secretary of War Pete Hegseth demanded that Anthropic remove its restrictions and permit what the Pentagon framed as “all lawful uses” of the technology, without specifying any ethical boundaries the military would not cross. Anthropic CEO Dario Amodei refused, arguing that “AI-driven mass surveillance presents serious, novel risks to our fundamental liberties,” and that current AI systems are not reliable enough to be trusted with fully autonomous weapons.

President Trump responded on Friday, Feb. 28, by ordering all federal agencies to “immediately cease all use of Anthropic’s technology.” Hegseth followed by designating Anthropic a supply chain risk to national security, a classification previously reserved for foreign adversaries, effectively barring any military contractor or supplier from doing business with the company. Hours later, OpenAI announced its own Pentagon deal under terms that, by contrast, carry no specified ethical limits beyond legality.

The public reaction to that deal was swift and measurable. According to market intelligence firm Sensor Tower, US uninstalls of the ChatGPT mobile app surged 295% day-over-day on February 28, compared with a typical daily uninstall rate of around 9%. Meanwhile, US downloads of Claude jumped 51% on the same day, and the app reached number one on the U.S. App Store by Saturday, a climb of more than 20 ranks in under a week. 1-star reviews for ChatGPT surged 775% on the same day, while 5-star reviews fell by half. In other words, a segment of the public was registering through consumer behaviour a verdict that the political system had not yet been asked to deliver.

The broader significance of this sequence reaches well beyond a contract dispute. In 2020, the Department of Defense announced its own AI principles, requiring that military AI be responsible, equitable, traceable, reliable, and governable. NATO followed with similar principles in 2021, and the United Kingdom in 2022. Those frameworks sent a clear signal to Russia, China, and every other major military power about how AI in warfare should be governed.

By sidelining the AI company insisting on those principles, and replacing it with a supplier willing to accept no restrictions beyond legality, the Trump administration effectively withdrew from the international consensus it had helped to build. The phrase “all lawful purposes” carries the appearance of a guardrail while offering none of the substance of one, given that laws change, interpretations shift with administrations, and governments define legality to suit their own priorities. As Amodei himself argued, that language cannot provide stable protection against misuse of a technology whose capabilities evolve faster than legislation can track.

The administration’s political framing further obscured what was at stake. By labelling Anthropic a “radical left, woke company” run by “leftwing nut jobs,” Trump and Hegseth mapped a fundamentally technical and legal dispute onto a culture-war framework. In reality, the backlash against Hegseth’s actions came from across the political spectrum, including from figures who had authored the administration’s own AI policy and who argued the move struck at core principles of private property.

Moreover, the administration’s own behaviour undermined its stated rationale. Even as the ban took effect, the U.S. military continued using Claude throughout the Iran operation, because the model was so deeply embedded in operational systems that removing it on short notice was operationally impossible. The Pentagon was simultaneously blacklisting and depending upon the same technology, a contradiction that showed more vividly than any policy document how indispensable AI had become to American warfighting.

What makes this governance crisis particularly consequential is the absence of any disclosure obligation. It remains publicly unknown what, precisely, Claude did in the Iran strikes. Did it flag specific buildings? Estimate casualty numbers? Recommend strike sequences? No official has answered these questions, and no law requires them to. A useful historical parallel is the U.S. drone program launched after Sept. 11, 2001. The government refused to acknowledge that program for years, and it took 15 years of lawsuits and sustained public pressure before the Obama White House published even basic casualty figures in 2016, figures that were widely regarded as undercounts. AI in warfare will be harder still to police, given that its outputs are less visible and its operational role more deeply embedded.

The Iran conflict marks a threshold that is still being crossed. AI has moved from experimental tool to load-bearing operational infrastructure, and it has done so faster than any governance framework can track. The debate that erupted in February 2026 focuses, understandably, on the confrontation between Anthropic and the Pentagon. However, that debate risks missing the larger and more urgent question, which is not whether AI should be used in warfare, but under what rules, with what transparency, and with what accountability it is being used right now.

Every military on earth is watching what is happening over Iranian skies and drawing the same conclusion, namely that AI-assisted warfare is effective and that the world’s most powerful military is demonstrating that publicly in real time. Therefore, the international community faces an urgent and shrinking window to establish binding constraints, transparent procurement standards, and independent oversight before the normalisation of AI in military operations becomes so routine that no one thinks to ask anymore.

Baggiarini, Bianca. 2026. “The Pentagon Strongarmed AI Firms before Iran Strikes – in Dark News for the Future of ‘Ethical AI.’” Edited by Michael Lucy. The Conversation , March. https://doi.org/10.64628/aa.4kdscuv4m.

Booth, Robert, and Dan Milmo. 2026. “Iran War Heralds Era of AI-Powered Bombing Quicker than ‘Speed of Thought.’” The Guardian. The Guardian. March 3, 2026. https://www.theguardian.com/technology/2026/mar/03/iran-war-heralds-era-of-ai-powered-bombing-quicker-than-speed-of-thought.

Brown, Larisa. 2026. “How AI Helps 20 US Troops Do the Work of 2,000 in Iran War.” Thetimes.com. The Times. March 3, 2026. https://www.thetimes.com/world/middle-east/article/palantir-ai-software-us-iran-war-lwld892z9.

Mchugh, Calder . 2026. “The War with Iran Is Scrambling the Politics of AI.” POLITICO. March 2, 2026. https://www.politico.com/newsletters/forecast/2026/03/02/the-war-with-iran-is-scrambling-the-politics-of-ai-00807434.

Olson, Parmy. 2026. “How Anthropic’s Claude AI Helped US Bomb Iran.” Www.ndtv.com. NDTV. March 4, 2026. https://www.ndtv.com/opinion/iran-israel-us-war-how-anthropics-claude-ai-helped-us-bomb-iran-11166940.

Perez, Sarah. 2026. “ChatGPT Uninstalls Surged by 295% after DoD Deal | TechCrunch.” TechCrunch. March 3, 2026. https://techcrunch.com/2026/03/02/chatgpt-uninstalls-surged-by-295-after-dod-deal/.

Rowland Manthorpe. 2026. “AI Could Be Giving US Lethal Edge in Iran War – but There Are Dangers.” Sky News. Sky. March 3, 2026. https://news.sky.com/story/ai-could-be-giving-us-lethal-edge-in-iran-war-but-there-are-dangers-13514784.

Steinberg, Jon, and Ezra Wallach. 2026. “War in Iran, Crisis in AI: SF Reacts to Combat, Divergent Moves by Anthropic, OpenAI.” Sfstandard.com. The San Francisco Standard. February 28, 2026. https://sfstandard.com/2026/02/28/war-iran-crisis-ai-sf-reacts-war-diverging-moves-anthropic-openai/.

Stokel-Walker, Chris. 2026. “Trump Is Using AI to Fight His Wars – This Is a Dangerous Turning Point.” The Guardian. The Guardian. March 3, 2026. https://www.theguardian.com/commentisfree/2026/mar/03/trump-using-ai-to-fight-wars-dangerous-us-military.

Vaast, Emmanuelle. 2026. “From Anthropic to Iran: Who Sets the Limits on AI’s Use in War and Surveillance?” Edited by Heather Walmsley. The Conversation, March. https://doi.org/10.64628/aam.fjcnwkwgn.