Ukraine: The First AI War as Living Laboratory

When Russia launched its full-scale invasion in February 2022, few analysts anticipated that the conflict would rapidly become the most consequential proving ground for AI-assisted combat since the advent of precision-guided munitions. Ukraine, fighting for its survival against a larger adversary, had both the motive and the pre-existing talent base to innovate. With over 240 AI-focused companies and a commercial sector that had supplied voice synthesis to Disney and navigation algorithms to Qualcomm, the country brought considerable civilian AI capacity to a military footing almost overnight.

The results were striking. By retraining publicly available AI models on real battlefield data, Ukrainian engineers pushed drone strike accuracy from roughly 10 to 20 percent to around 70 to 80 percent, a three- to four-fold improvement that transformed what had been primarily reconnaissance platforms into precision strike weapons. The country produced 800,000 drones in 2023 and doubled that figure in 2024, with Ukrainian officials targeting 5 million units in 2025, at least half equipped with AI guidance. Beyond drones, AI played a foundational role in intelligence fusion: neural networks combined satellite imagery, open-source social media data, and ground-level video to geolocate Russian forces, while Clearview AI’s facial recognition was used to identify fallen soldiers and expose disinformation.

Throughout this period, Ukraine maintained what analysts called a human-centric approach, with a human operator retaining final targeting authority even as AI dramatically compressed the time between identification and engagement. Ukraine’s Delta system, upgraded in 2024 with machine-learning capabilities for real-time video and text processing, became the nerve center of this approach, fusing intelligence streams and presenting targeting options to commanders who still held the trigger. The ethical architecture of the Ukrainian model was one of AI as accelerant, not as autonomous decision-maker.

Venezuela: AI Enters the Classified Operations Domain

The operation that removed Maduro from power on January 3, 2026, marked a qualitative leap in the military use of AI, not because of the scale of its application, but because of its domain. The Wall Street Journal’s subsequent reporting, confirmed by multiple sources cited by Axios, revealed that the U.S. military had used Anthropic’s Claude language model during the active phase of the raid, not merely in preparatory planning. Accessed through Palantir Technologies’ defense contracting infrastructure, Claude became the first commercially developed large language model known to have been deployed in a classified U.S. military operation.

The precise role Claude played remains deliberately opaque, as neither the Pentagon nor Anthropic has confirmed specifics, but reporting indicates that AI capabilities were used to analyze satellite imagery, process intelligence, and potentially assist in targeting the multiple sites struck across Caracas. Venezuela’s Defense Minister Vladimir Padrino López subsequently claimed that the United States had used the country as a weapons laboratory for technologies powered by artificial intelligence that had never before been deployed on a battlefield, accusations that dovetailed with President Trump’s own cryptic references to a device he called the “discombobulator”, said to have disabled Venezuelan defensive systems.

The operation also revealed the complex governance fault lines that AI militarization was beginning to expose. Anthropic’s usage policies explicitly prohibited Claude from facilitating violence, weapons development, or surveillance. The Pentagon, for its part, insisted that it should be able to deploy commercial AI however it chose, provided it complied with U.S. law. The friction generated by the Maduro raid would prove consequential, foreshadowing a rupture between Anthropic and the Department of Defense that would come to a head only weeks later. Meanwhile, AI-generated disinformation swirled around the operation almost immediately, with fabricated videos purporting to show joyful Venezuelans racking up millions of views before fact-checkers could respond, illustrating that the information dimension of AI warfare was now as significant as its kinetic dimension.

Iran: AI at Industrial Scale

If Ukraine was the laboratory and Venezuela the proof of concept, the Iran campaign has been the mass deployment. In the first 24 hours of U.S. strikes beginning February 28, 2026, the Pentagon struck roughly 1,000 targets, a pace that would have been operationally impossible without AI-assisted target generation and prioritization. The Washington Post and subsequent reporting in Nature confirmed that Claude, operating through the Pentagon’s Maven Smart System and Palantir’s classified infrastructure, had proposed hundreds of targets, ranked them by priority, and supplied precision coordinates derived from satellite and signals intelligence, all before human commanders gave final approval for each strike.

The Maven Smart System, originally developed for image processing and tactical intelligence in earlier Middle East operations, had by this point been deeply integrated with large language model capabilities. Israel’s parallel contribution layered its own AI infrastructure onto the campaign: synthesizing traffic camera footage hacked from Iranian networks, billions of data points, and the AI-assisted targeting system known as Lavender, previously used in Gaza, to build a targeting bank of thousands of Iranian sites. Israel also used AI-generated content as a psychological warfare instrument, hacking a widely used Muslim prayer application to send defection messages to Iranian soldiers and seeding AI-generated imagery of prison bombings to Iranian officials and the public.

The Iran theater represented an effective removal of the human from meaningful deliberation in parts of the kill chain. According to investigators at Responsible Statecraft and accounts from defense analysts cited by Nature, the volume and speed of AI-recommended strikes made genuine human review of each target a practical fiction. Defense Secretary Pete Hegseth’s declaration that there would be no stupid rules of engagement, alongside his earlier assertion that the future of American warfare is spelled AI, signaled a policy posture that prioritized operational tempo over independent verification. The consequences became visible almost immediately: the obliteration of the Shajareh Tayyebeh girls’ school in Minab, killing at least 165 students and staff, raised questions that the Pentagon declined to answer when asked whether AI had been involved in selecting that target.

Table 1 below summarizes the main dimensions of AI use in the three warfare theaters explored above.

Table 1. AI in Warfare Across Three Military Theaters (2022-2026)

The next Similarities Across the Three Theaters

Several throughlines connect all three operations. In each, AI served primarily as an intelligence fusion and targeting acceleration tool, compressing what military planners call the kill chain by synthesizing data faster than human analysts could and presenting actionable options at machine speed. In each theater, commercial AI platforms and their underlying companies were pulled into military applications for which they were not designed, generating governance tensions between profit, safety commitments, and state power. And in all three cases, the information environment surrounding kinetic operations was saturated with AI-generated content, both as a tool of the combatants and as an unpredictable byproduct exploited by domestic political actors.

There is also a consistent underlying ethical ambiguity: none of the three theaters has resolved the question of meaningful human control. Ukraine came closest to a defensible answer, but even there the speed of AI-assisted drone strikes raised concerns about how much deliberation was actually occurring at the moment of engagement. In Venezuela and Iran, the claim of human authorization has become increasingly formulaic, a legal cover story for what are effectively algorithmic decisions executed at a pace that forecloses genuine oversight.

Evolution: From Tool to Architect

The arc from Ukraine to Iran traces a clear evolutionary trajectory. In Ukraine, AI was a battlefield accelerant integrated into a deliberate, iterative system with trained human operators making final calls. The emphasis was on improving existing processes: better drone guidance, faster intelligence analysis, sharper geolocation. The innovation was impressive, but the human remained notionally at the center.

In Venezuela, the threshold shifted. A commercial large language model entered a classified real-time operation, no longer merely supporting planning, but embedded in the execution itself. The operation was still small enough, and the targets precise enough, that oversight remained at least partially plausible. But the institutional guardrails began to bend: a company whose terms of service prohibit facilitating violence found its technology used in a raid that killed over 80 people, and the primary response from the Pentagon was to threaten the company’s contract for asking questions.

In Iran, the architecture of restraint largely collapsed. AI moved from augmenting human targeting decisions to effectively generating them at a scale and speed that rendered human review nominal. The Maven Smart System, powered by Claude and supported by allied AI platforms, became less a decision-support tool and more a decision-making system with human countersignatures attached for legal compliance. The targeting error at Minab may or may not have involved AI, but the structure of the operation made such errors predictable: when machines propose a thousand targets in a day, the statistical probability of algorithmic error producing catastrophic civilian harm approaches certainty.

What has evolved, then, is not simply the technology, but the political and institutional willingness to subordinate human judgment to machine speed. Ukraine pioneered the integration of AI into warfare under the pressure of existential threat and maintained, however imperfectly, a commitment to operator accountability. The United States, operating from a position of overwhelming superiority and against adversaries of its own choosing, has progressively dismantled that commitment. The three theaters together suggest that AI in warfare follows a path not unlike nuclear weapons in the Cold War: once the capability exists and its operational advantages are demonstrated, the logic of escalation tends to overwhelm the logic of restraint. The question the Iran campaign poses most urgently is whether there remains a political constituency, inside or outside the government, capable of reimposing meaningful limits before theater strips them away entirely.

REFERENCES

Ukraine

• War Room (U.S. Army War College) — “Artificial Intelligence’s Growing Role in Modern Warfare” (August 2025)

• CNAS / Russia Matters — “Roles and Implications of AI in the Russian-Ukrainian Conflict” (July 2023)

• Second Line of Defense — “How Ukraine Transformed Western Military Aid and Domestic Innovation into a Dynamic Defense Model” (August 2025)

• Atlantic Council — “Missiles, AI, and Drone Swarms: Ukraine’s 2025 Defense Tech Priorities” (January 2025)

• AI Frontiers — “How AI Is Eroding the Norms of War” (May 2025)

• CSIS — “Understanding the Military AI Ecosystem of Ukraine” (January 2025)

• Spanish Ministry of Defense (DIEEA) — “Artificial Intelligence and the War in Ukraine” (2024)

• Friends of Europe — “The AI Frontier: Ukraine’s Role in the Future of Warfare” (October 2024)

• Taylor & Francis — “Navigating the AI Frontier: NATO’s Governance Role in Military AI” (February 2025)

Venezuela

• Axios — “Pentagon used Anthropic’s Claude during Maduro raid” (February 13, 2026)

• Fox News Digital — “AI tool Claude helped capture Venezuelan dictator Maduro” (February 2026)

• TRT World — “US military used Anthropic’s Claude AI in Maduro abduction raid” (February 2026)

• Business and Human Rights Centre — “US military allegedly used Anthropic AI in Venezuelan president’s abduction” (February 2026)

• American Bazaar — “US military used Claude AI in Venezuela to capture Maduro” (February 16, 2026)

• Politics Today — “Venezuela Accuses U.S. of Using ‘Secret Weapon’ in Maduro Abduction” (January 27, 2026)

• Al Jazeera — “The ‘discombobulator’: Did US use ‘secret weapon’ in Maduro abduction?” (January 26, 2026)

• Futurism — “That Video of Happy Crying Venezuelans After Maduro’s Kidnapping? It’s AI Slop” (January 5, 2026)

• CEDMO — “AI, outdated visuals fuel misinformation after Maduro capture” (January 9, 2026)

• Small Wars Journal — “AI-Enabled Decapitation Strike: Maduro Raid” (February 17, 2026)

Iran

• Washington Post — “Pentagon leverages AI in Iran strikes amid feud with Anthropic” (March 4, 2026)

• Nature — “How AI is shaping the war in Iran” (March 5, 2026)

• CNN — “Pentagon using AI in Iran strikes as tech firms clash over military use” (March 7, 2026)

• Japan Times — “Use of AI for Iran targeting raises questions amid growing military reliance” (March 9, 2026)

• NPR — “Old-school tricks and AI tech are weapons in the Iran war” (March 10, 2026)

• Bloomberg — “Iran War Provides a Large-Scale Test for AI-Assisted Warfare” (March 5, 2026)

• Responsible Statecraft — “US used ‘Claude’ to strike over 1,000 targets in first 24 hours of war” (March 2026)

• South China Morning Post — “Iran strikes are a wake-up call to regulate military AI” (March 10, 2026)

• MIT Technology Review — “How AI is turning the Iran conflict into theater” (March 9, 2026)

• Futurism — “Pentagon Refuses to Say If AI Was Used to Select Elementary School as Bombing Target” (March 2026)