Project Maven: The Book That Exposed America's AI War Machine

On the morning of March 26, 2026, two of America's most influential business and technology news organizations published major features on the same subject at the same hour. NPR and Bloomberg both covered the release of journalist Katrina Manson's new book, simply titled Project Maven. The simultaneity was not accidental. The book — a meticulous, year-in-the-reporting account of how the United States military industrialized artificial intelligence into its targeting and kill chain architecture — arrives as the Iran conflict provides its first real-world confirmation that the system works exactly as designed. Or, depending on your vantage point, exactly as feared.

This is not a story about a technology demonstration. It is a story about institutional transformation at the speed of software. It is the story of how a small, classified Pentagon initiative launched nine years ago without public consent, without international treaty, and without meaningful civilian oversight became the cognitive engine of American warfare. And it is the story of the people — engineers, protesters, generals, executives, and ethicists — who either built it, tried to stop it, or quietly enabled it.

The Origin: 2017 and the Data Overload Problem

The problem Project Maven was created to solve was, on its surface, mundanely administrative. By 2017, the US military's fleet of drones — Predators, Reapers, Global Hawks — was generating more raw video footage per day than any human team could possibly review. Analysts were drowning. The ratio of sensor data to analytical bandwidth had inverted catastrophically. Drones could see everything; the military lacked the cognitive throughput to understand what it was seeing in anything approaching real time.

Deputy Secretary of Defense Robert Work issued a directive in April 2017 formally establishing the Algorithmic Warfare Cross-Functional Team. It was immediately given an informal name that stuck: Project Maven. Its stated mission was to accelerate the integration of artificial intelligence and machine learning into the Department of Defense. Its immediate practical task was simpler: build a computer vision system that could watch drone footage and identify objects of military interest. Vehicles. People. Weapons. Patterns of movement. The system would not fire weapons. It would not make kill decisions. It would, in the department's framing, "augment" human analysts by surfacing relevant data faster.

The Pentagon needed a technology partner quickly. It turned to the private sector. The initial contract went to Google.

The Google Revolt: When Engineers Said No

The partnership between Google and the Pentagon on Project Maven was not announced. It was discovered. In early 2018, internal communications began circulating within Google's engineering and research departments. Employees learned that the company had signed a contract with the Department of Defense to provide TensorFlow AI technology for drone video analysis. The reaction was immediate and forceful.

"We believe that Google should not be in the business of war."

— Open letter signed by more than 4,000 Google employees, April 2018

The petition that circulated internally gathered over four thousand signatures from employees across Google's engineering, product, and research divisions. A smaller number — estimated at a dozen or more — resigned outright in protest. The signatories were not fringe voices. They included senior engineers and respected researchers whose contributions underpinned Google's core products. Their argument was precise: providing AI capabilities for weapons targeting crossed an ethical line that no contract, however lucrative, could justify. They demanded that Google establish clear policy never to build warfare technology or weapons systems.

The external pressure arrived in parallel. Academic researchers, ethicists, and civil liberties organizations began scrutinizing the contract. Op-eds appeared in the New York Times and the Guardian. The story became a referendum on the technology industry's relationship with the national security state — a relationship that had been largely invisible to the public since the Snowden revelations of 2013.

In June 2018, Google announced it would not renew its Project Maven contract when it expired. The company simultaneously released its AI Principles, a set of ethical guidelines that included an explicit commitment not to build AI weapons systems. It was a significant symbolic victory for the employee protesters. It was also, from a military program management perspective, a temporary inconvenience.

Who Stepped In: Palantir and the Permanent Platform

The contract that Google walked away from was not left unfilled for long. Palantir Technologies, founded in 2003 with seed funding from the CIA's venture arm In-Q-Tel and built from the ground up to serve intelligence and defense customers, stepped into the gap. Palantir had no employee ethics petitions to worry about. Its workforce understood from day one that the company existed to serve exactly the kind of mission that had discomfited Google's engineers. National security was not a sidebar to Palantir's business — it was the business.

Under Palantir's stewardship, Project Maven did not merely continue. It was architecturally redesigned. The original system had been a relatively narrow computer vision tool — sophisticated, but bounded. Palantir's approach, built on its Gotham and Foundry platforms and later its AI Platform (AIP), was to transform Maven from a single-function analytical tool into a full-spectrum intelligence fusion environment. The new architecture could ingest and correlate multiple data streams simultaneously: drone video, satellite imagery, signals intelligence, communications intercepts, open-source data, and battlefield sensor networks. It could identify not just objects but patterns of life — the regularized behaviors of individuals and groups over time that, in targeting doctrine, constitute the intelligence basis for lethal action.

Palantir's leadership was explicit about its ambitions. CEO Alex Karp described the company's mission in nearly martial terms in public statements and shareholder communications. Where Google had retreated from defense AI amid internal protest, Palantir advanced. As we reported in March 2026, the Pentagon moved to formalize Project Maven as a permanent long-term acquisition program — the bureaucratic designation that signals an intent for decades of continued investment rather than experimental cycles.

The transformation from Google skunkworks experiment to permanent Palantir platform took less than eight years. In institutional terms, that is extraordinarily fast.

The Evolution: From Image Analysis to Kill Chain Acceleration

To understand what Project Maven became, it helps to understand the concept of the kill chain — the sequence of activities required to move from target identification to lethal action. The classic formulation, known as F2T2EA, runs: Find, Fix, Track, Target, Engage, Assess. Each step has historically required human analysis, human communication, and human decision-making. The process could take hours, days, or longer. Against fast-moving targets, that latency was operationally catastrophic. Adversaries could disappear between the "find" and "engage" steps.

Project Maven's evolution has been, in fundamental terms, an effort to compress the kill chain. Not to remove humans from it — current Pentagon doctrine and the DoD's Directive 3000.09 formally require a human being to authorize lethal action — but to reduce the cognitive and temporal burden on the human decision-makers at each step. The AI does not pull the trigger. The AI does everything else at machine speed, presenting the human operator with a compressed, pre-analyzed targeting recommendation accompanied by a confidence score and supporting intelligence data.

By 2025, public reporting and defense contractor disclosures indicated that Maven-derived systems were performing the following functions: automated detection and classification of targets from persistent surveillance feeds, pattern-of-life analysis aggregating months of behavioral data, autonomous tracking of designated targets across multiple sensor inputs, automated correlation of signals intelligence with imagery to confirm target identity, and pre-populated targeting packages ready for command review. The system had expanded far beyond drone video. It was now the connective tissue between the entire intelligence collection architecture and the targeting authorization process.

The Iran Deployment: AI Goes to War

The confirmation arrived in stages, as military confirmations usually do. In late February 2026, multiple defense reporters began noting unusual specificity in the speed of American targeting operations against Iranian and Iranian-affiliated assets. Strike cycles that had historically taken 12 to 24 hours of analytical preparation were completing in under two hours. The target packages were described by military officials, in background conversations, as the most precisely developed in American combat history.

By early March 2026, the Defense Department acknowledged — without specific technical detail — that AI-assisted intelligence analysis was supporting operations. The acknowledgment was understated. What the department described as "AI-assisted" was, in the characterization of multiple defense analysts and former officials speaking to Bloomberg and NPR, a system in which AI-generated targeting recommendations were driving the operational tempo. Human authorization remained the formal final step. The humans were working from AI-constructed packages at a pace that precluded meaningful independent verification.

This is the operational reality that Katrina Manson's book documents. The Iran deployment is not an anomaly or an emergency exception. It is the designed endpoint of a nine-year development program. Maven was always being built to do exactly this. The question was never whether it would reach this capability — it was whether anyone would stop it before it did.

Manson's Revelations: What the Book Exposes

Katrina Manson spent the better part of three years reporting the book that shares its title with the program it documents. A former Financial Times defense correspondent who covered the Pentagon, Manson had access that most journalists pursuing this subject do not. She interviewed former and current Defense Department officials, AI researchers who worked on the program, technology company executives, military ethicists, and intelligence officers. The result is a level of documentary specificity that makes the book uncomfortable reading for anyone who assumed that the transition to AI-assisted warfare was subject to meaningful oversight.

The book's central revelations cluster around several themes:

The Oversight Gap

Manson documents in detail the structural gap between the pace of AI capability deployment and the pace of policy and oversight development. The military's acquisition processes, designed for hardware procurement cycles measured in decades, were not built to govern software systems that could be updated, retrained, and fundamentally altered between oversight reviews. By the time Inspector General reports or congressional hearings could examine a given iteration of Maven's capabilities, the system had already moved several generations forward. The oversight apparatus was perpetually reviewing a past version of the technology.

Internal documents obtained by Manson show that multiple DoD ethics review processes flagged concerns about the speed of Maven's expansion into targeting functions beyond its original image analysis mandate. Those concerns were noted, documented, and — in the operational urgency created by emerging threats — consistently deferred. The bureaucratic record contains a remarkable artifact: a 2023 internal memo acknowledging that Maven's use for targeting recommendation functions had "outpaced the policy framework" governing its deployment. The memo recommended a review. The review was never completed.

Civilian Casualty Accounting

The book's most controversial chapter addresses the question of civilian casualties in AI-assisted strike operations. Manson draws on declassified after-action reports, comparative analysis with non-AI-assisted strike campaigns, and interviews with military lawyers and targeting officers to argue that the speed of AI-assisted kill chains systematically disadvantages the precautionary analysis required by international humanitarian law. The law requires that before a strike is authorized, a commander must take all feasible precautions to assess and minimize civilian harm. When targeting packages arrive pre-loaded with AI-generated confidence scores, that precautionary analysis is compressed into a binary: approve or reject the AI's work.

The data Manson presents is disputed by the Pentagon, but it tracks with findings published by independent analysts: in operations where AI-assisted targeting was used, civilian casualty incidents were not eliminated. They occurred at rates that, in several documented cases, correlated with lower-confidence AI classifications that human analysts, given time and independent intelligence, might have flagged for additional review. The AI was not making wrong calls with unusual frequency. But the speed of the process meant that when it was wrong, the error was acted upon before it could be caught.

Corporate Acceleration

Manson reserves some of her most pointed analysis for the commercial incentive structures that shaped Maven's evolution. The competition for defense AI contracts created a market dynamic in which vendors had financial incentives to demonstrate expanded capabilities on compressed timelines. Palantir, Anduril, Shield AI, and a growing ecosystem of smaller companies competed for program dollars by demonstrating increasingly sophisticated targeting and kill chain functions. The Pentagon's procurement officers, under pressure to show progress and capability, rewarded demonstrated functionality over process maturity. The result, as Manson frames it, was a system optimized for capability demonstration rather than reliability, accountability, or civilian safety.

Wartime Speed: The Hegseth Directive

The political context for Maven's operational deployment in Iran was established in January 2026, when Defense Secretary Pete Hegseth issued what became known internally as the "wartime speed" directive. The document — a memo distributed to senior Pentagon leadership and defense contractors — instructed the department to eliminate bureaucratic friction from AI capability deployment in the service of operational readiness. Specifically, it directed that AI systems demonstrating "operational utility" could proceed to deployment pending ongoing, rather than completed, safety and ethics review processes.

The directive was consistent with Hegseth's stated philosophy of running the Defense Department like a wartime institution even in the absence of formally declared war. It was also, in the interpretation of several former DoD lawyers and policy officials who spoke to NPR on background following the book's publication, a formal inversion of the department's own AI ethics framework. The 2020 DoD AI Ethics Principles require that AI systems be reliable, governable, traceable, and subject to human oversight before operational deployment. The wartime speed directive effectively subordinated those requirements to operational need.

"We are not going to wait for perfect when adversaries are deploying capable. Speed is a force multiplier. The bureaucracy that slows us down is itself a national security threat."

— Internal DoD memo attributed to Hegseth directive, January 2026, as cited in Manson's book

The directive did not generate significant public attention when it was issued. It generated rather more when, two months later, its consequences became visible in the Iran operation's targeting tempo. The connection — between a January policy memo and March operational reality — is one of the threads Manson traces with particular care.

The Ethical Vacuum: No Treaty, No Framework, No Brake

The international context for Maven's operational deployment is bleak. As of March 2026, there is no international treaty governing the use of artificial intelligence in warfare. There is no binding international law specifically addressing autonomous targeting systems. There are no agreed international standards for the level of human oversight required before AI-generated targeting recommendations can be acted upon. There is no international verification regime, no inspection mechanism, no enforcement body.

The Convention on Certain Conventional Weapons has been the primary international forum for discussions of Lethal Autonomous Weapons Systems since 2014. Twelve years of expert group meetings, informal consultations, and formal negotiating sessions have produced guidelines, principles, and recommendations — none of which are legally binding, and most of which the United States, Russia, and China have not endorsed in their strongest forms. See our full coverage of the CCW process and the international ethics framework.

The United States' position, consistently maintained across administrations, is that existing international humanitarian law is sufficient to govern AI-enabled warfare, and that new treaty frameworks are premature given the pace of technological change. The argument has a certain coherence: IHL principles of distinction, proportionality, and precaution apply to any weapons system regardless of the technology involved. The problem, as Manson and a growing body of legal scholars argue, is that these principles were designed for decision timelines measured in human perception — seconds to minutes. Applied to AI systems operating at machine speed, with targeting recommendations generated faster than human cognition can meaningfully evaluate them, the principles become procedural fictions. The law says a human must be in the loop. A human is technically present at every step. But the human is running at 3,000 feet per second through a targeting process that was built to go at Mach 2.

The policy landscape reflects this gap. The doctrine has not kept pace with the capability. And the Iran conflict is the first live test of what that gap means in practice.

What Comes Next: AI as Permanent Warfare Infrastructure

Manson's book closes with a series of observations that function less as predictions than as documentation of trajectories already established. Several are worth examining carefully, because they describe a future that is already arriving.

Normalization and Precedent

The Iran deployment has established a precedent that will be difficult to reverse. The United States has now demonstrated, in live combat operations, that AI-assisted targeting is operationally viable, that it compresses the kill chain to advantage, and that the political and legal objections to its use do not constitute an operational barrier. Future administrations, facing future adversaries, will have both the capability and the precedent to deploy similar systems without the deliberation that even the current deployment — arguably inadequate — received. The first use establishes the template for all subsequent uses.

Adversary Response

China and Russia have been developing AI targeting capabilities in parallel with the United States for years. Both have conducted their own versions of Google's internal debate — and resolved it, as Palantir did, in favor of capability development. China's military AI programs, documented in PLA research and defense white papers, explicitly target the same kill chain compression that Maven pursues. Russia's experience in Ukraine has served as a live testing environment for AI-enabled drone operations. The Iran deployment, however it is evaluated in terms of legal or ethical compliance, is likely to accelerate adversary AI weapons development. The competitive dynamic is not deterred by American capability demonstration. It is amplified by it.

The Contractor Ecosystem

One of Manson's most important structural observations concerns the private sector ecosystem that has grown up around defense AI. Palantir is not alone. Anduril, founded by Palmer Luckey with explicit mission alignment to defense AI and autonomous systems, is building the hardware and software infrastructure for what it describes as the autonomous military of the future. Shield AI builds autonomous fighter and drone systems. Rebellion Defense, acquired by Shield AI, built software for exactly the kind of intelligence fusion Maven pioneered. A dozen smaller companies fill in the gaps. The collective revenue of this ecosystem now runs into the tens of billions of dollars annually, and it is growing at rates that track defense budget allocation rather than commercial market cycles. The contractors have strong financial incentives for Maven to expand, for its deployment envelope to widen, and for the policy frameworks that might constrain it to remain permissive. Those incentives do not make ethical failure inevitable. But they make it systematically more likely.

The Autonomy Gradient

The final trajectory Manson traces is the most technically specific and, arguably, the most consequential. The current architecture of Project Maven maintains human authorization as the final step in the targeting chain. That authorization is increasingly pro forma given the speed and volume of AI-generated packages, but it is structurally present. Future iterations of the system, already in development according to sources cited in the book, are designed for environments where communication latency or electronic warfare makes real-time human authorization impossible — contested airspace, submarine operations, cyber operations with millisecond decision windows. For those environments, the architecture must, by operational necessity, push the human further from the decision point. The gradient from "human in the loop" to "human on the loop" to "human out of the loop" is not hypothetical. It is a planned engineering progression, with each step justified by operational requirements that are real.

Where to Go From Here

Project Maven does not exist in isolation. It is the visible center of a much larger transformation in how artificial intelligence is being integrated into every aspect of American — and global — military capability. Our coverage of this transformation spans the technology, the policy, the doctrine, and the specific conflicts where these systems are now operating.

• For detailed coverage of Palantir's role and Maven's permanent program status, see our earlier report: Palantir's Maven Goes Permanent: The AI That Sees Everything
• For the Iran conflict context and how AI targeting is operating in the current war, see our Iran War coverage
• For the doctrine framework — how the military conceptually governs AI use in operations — see our Doctrine section
• For the policy and legal frameworks, including international humanitarian law, the CCW process, and DoD directives, see our Policy section
• For the full ethical and legal analysis, including the human-in-the-loop spectrum and international law failures, see our AI Ethics and International Law page

The story of Project Maven is the story of the most consequential technological shift in the history of warfare. It is happening now. It is not finished. And its implications — legal, ethical, strategic, and human — are still being written.