In the time it takes you to read this sentence, an artificial intelligence system designed by the Pentagon can identify a person, classify them as a target, and route a missile to their location. A human operator still has to press the button. But the button-press is not what we usually mean by "judgment." It is a reflex — a twitch of the thumb at the end of a process that began in a server farm and raced through 179 data feeds faster than any human could follow.
This is Project Maven. It started as a homework assignment for computers: look at drone footage and tell us if that blob is a truck or a tent. Nine years later, it is the central nervous system of American AI-enabled warfare — a system that processes up to 5,000 targets per day, that the Pentagon insists operates under "appropriate levels of human judgment," and that, in at least one documented case, helped kill more than 110 children before anyone noticed the algorithm had made a mistake.
The question is not whether Project Maven works. It works terrifyingly well. The question is whether the story the Pentagon tells about it — that a "human in the loop" guarantees moral accountability — holds up to scrutiny. The evidence suggests it does not. But let us walk through that evidence carefully, because the gap between marketing and reality is where skepticism lives.
What Project Maven Actually Is
Project Maven is the informal name for the Algorithmic Warfare Cross-Functional Team, created by Deputy Secretary of Defense Robert O. Work on April 26, 2017 (Department of Defense). Work was worried that China was pulling ahead in military AI, and he wanted the Pentagon to move fast. The project was given $70 million and a simple mandate: find ways to use machine learning for intelligence work.
At first, that meant teaching computers to sort through full-motion video from drones. Human analysts were drowning in footage. Colonel Drew Cukor, the project's first leader, explained that Maven would help identify 38 categories of objects — vehicles, buildings, weapons, people — so humans would not have to stare at screens for twelve hours straight (Shanahan). By December 2017, the team had manually labeled 150,000 training images. The pitch was modest: computers detect, humans decide.
Nothing about that initial description suggested weapons. And yet, even in 2017, Cukor described the goal as helping "weapon systems to detect objects" (Shanahan). It is a small semantic slide — from "help analysts" to "help weapon systems" — but it previews a pattern we will see again and again. Each step in Maven's expansion was framed as a natural technical evolution, instead of a deliberate policy choice.
From Footage Sorter to Kill Chain
Maven's growth can be tracked in four phases.
Phase One (2017-18) was computer vision. The system learned to recognize objects in drone video. That was it. A fancy filing system for spy footage.
Phase Two (2020-22) introduced live-fire integration. The corps' first Scarlet Dragon exercise produced what was described as the first AI-enabled artillery strike in U.S. Army history. That initial "digital target pass" took 743 minutes. By 2024, after years of iteration, the same process took under one minute ("Project Maven").
The exercise in January–February 2023 was a later, more advanced iteration — the sixth in the series. Maven spotted a tank in satellite imagery, a human approved the target, and the system directed a HIMARS rocket-artillery system strike.
Phase Three (2023-24) made Maven an official Program of Record — a formal, budgeted military acquisition program — and transferred it to the National Geospatial-Intelligence Agency. It was folded into a larger Pentagon initiative called Combined Joint All-Domain Command and Control — military-speak for "connect everything to everything." At U.S. Central Command, Maven began fusing 179 separate data sources: drone video, radar, infrared sensors, cellphone geolocation, social media posts, satellite imagery, and cyber intercepts (Salah).
Phase Four (2025-26) added large language models — the same technology that powers ChatGPT — to digest all that data and write targeting summaries for human operators. The Pentagon calls this "accelerated decision-making." A senior targeting officer told reporters that with Maven, he could decide on 80 targets per hour, compared to 30 without it. The whole targeting cycle now runs in under 60 seconds ("Project Maven and the Architecture of AI Warfare").
For comparison: during the 2003 invasion of Iraq, a single targeting cell required roughly 2,000 staff. The 18th Airborne Corps now does comparable work with 20 people. That's not "assistance," it's replacement.
The Interface
The system operators see is called Maven Smart System. It displays yellow boxes around potential targets and blue boxes around friendly forces or no-strike zones. It tracks aircraft, ships, supply lines, and the locations of key personnel. When a human makes a decision, the system can transmit it directly to a weapon platform. Internal documents refer to "Maven ATR: automatic target recognition" (Salah).
Here is the key detail: of the six standard steps in military targeting — identify, locate, filter, prioritize, assign, and fire — Maven now performs four. The human is technically responsible for the last two: assigning a weapon and authorizing the shot. But by the time the human sees a target, the machine has already done the work of finding it, placing it, screening it, and ranking it. The human is not analyzing. The human is supervising an analysis that is already finished.
The question is not just what the interface shows, but who gets to define what "human oversight" means in the first place.
The Anthropic Rejection
In July 2025, the Pentagon awarded a $200 million prototype contract to Anthropic — one of several identical awards to AI companies, alongside Google, OpenAI, and xAI — to integrate its Claude language model into Maven. Anthropic's CEO, Dario Amodei, asked for two contractual guarantees: no mass surveillance of Americans, and no fully autonomous lethal weapons without meaningful human oversight.
The Pentagon refused. In March 2026, the Department of Defense designated Anthropic a "supply chain risk" and ordered its products phased out of military use within six months ("Anthropic-Pentagon Battle Shows How Big Tech Has Reversed Course on AI and War"). The message was hard to miss: the Pentagon would rather eject a vendor than accept a legally binding limit on autonomous lethality.
The refusal suggests that "meaningful human control" is operationally negotiable. An institution that treats safety conditions as supply-chain risks is not treating them as constraints.
The Speed Problem
DoD Directive 3000.09, the Pentagon's official policy on autonomous weapons, requires that commanders and operators exercise "appropriate levels of human judgment over the use of force" (Department of Defense, DoD Directive 3000.09). The directive was updated in January 2023. It does not define:
- What "appropriate" means.
- How much time counts as "judgment."
- Whether clicking "approve" on a machine-generated recommendation qualifies.
- Who is accountable when the recommendation is wrong.
The vagueness serves a purpose: it allows the Pentagon to claim compliance while compressing the human role into a split-second reflex.
The International Committee of the Red Cross defines lethal autonomous weapons as systems that can "search for, detect, identify, track, select, and attack targets" without human intervention. Maven technically does not meet this definition, because a human still presses the button. But cognitive psychology tells us what happens when humans supervise high-speed algorithms: automation bias. The more reliable a system seems, the less likely operators are to question it. When you are reviewing 80 targets per hour, you are not exercising moral judgment. You are doing quality control on an assembly line — and humans are bad at that.
The Minab School Attack
The cost of that structural failure became visible during the 2026 Iran war. On February 28, 2026, according to investigative reports, a Tomahawk missile struck Shajareh Tayyebeh elementary school in Minab, Iran. Over 150 people died, including more than 110 children. The strike was reportedly the result of a decade-old database error that Maven processed at machine speed. The algorithm either misidentified the school or failed to filter it from the target list. The human operator had seconds to review a recommendation synthesized from 179 data feeds. The human approved. The missile fired ("2026 Minab School Attack"; "The School the Algorithm Forgot").
Official accounts attributed the strike to human error rather than the algorithm. This is technically true in the narrowest sense: a human being clicked the button. But it is misleading in every way that matters. The human could not have verified the underlying data. The human did not know the algorithm was relying on stale information. The human was structurally incapable of providing the "appropriate judgment" the Pentagon claims is safeguard ("The School the Algorithm Forgot").
And then the system was promoted. According to the same reports, Maven was expanded — not paused, not independently audited, but scaled up — in the days following the Minab strike. The targeting rate, after the integration of large language models, increased from fewer than 100 targets per day to 1,000, and then to 5,000 ("Project Maven and the Architecture of AI Warfare").
Who Builds This?
The Minab strike was not caused by a rogue contractor. It was caused by a system whose accountability fragments across dozens of companies.
Google was Maven's first major AI partner, contributing TensorFlow, Google's open-source machine-learning toolkit, in 2017. When more than 3,000 Google employees signed an open letter declaring that "Google should not be in the business of war," the company walked away and published its "AI Principles" ("The Business of War"). Palantir took Google's place and now holds the primary contract, valued at up to $1.3 billion through 2029 ("Palmer Luckey's Anduril Won Project Maven Pentagon AI Contract"). Anduril, founded by Oculus Rift creator Palmer Luckey, supplies edge hardware and sensor networks. Booz Allen Hamilton received a $751.5 million prime award in 2018.* Enabled Intelligence holds a $708 million ceiling for data labeling. Amazon Web Services provides cloud infrastructure. At least 32 companies have worked on Maven as of March 2026.
The known financial commitments exceed $3.7 billion, and classified add-ons likely push the total higher. With so many contractors handling discrete stages — data ingestion, model training, cloud hosting, operator interface — no single node owns the full pipeline. That makes assigning responsibility after a failure a legal and bureaucratic labyrinth.
The Rhetorical Slide
Track how official descriptions of Maven have changed, and the pattern is unmistakable:
- 2017 - "People and computers will work symbiotically to increase the ability of weapon systems to detect objects."
- 2018 - "Human-in-the-loop decision support."
- 2022 - Became a formal Program of Record.
- 2024 - A "data-centric warfighting system that feeds apps for planning, preparing, and executing operations."
- 2025 - "Production-level across every combatant command" except Special Operations.
- 2026 - "Deploying across the entire department." Twenty-five thousand users.
Notice what never happened. There was no public debate. No congressional vote on whether an AI system that fuses 179 data sources and routes lethal targeting recommendations directly to weapons should become standard across the entire U.S. military. Each stage was presented as inevitable — the natural next step for a pathfinder project that was always supposed to kindle a larger flame.
The Verdict
So what does the evidence show?
It shows that Project Maven works as advertised: it dramatically accelerates targeting, compresses the kill chain to under a minute, and replaces thousands of human analysts with a small team supervising algorithmic output. The Pentagon's claims about "human in the loop" oversight are not false in the literal sense — a human does, physically, press the button. But they are misleading in the ways that matter for accountability and moral judgment.
The evidence shows that when a vendor tried to write meaningful human oversight into a contract, the Pentagon rejected the contract. It shows that when an algorithmic error contributed to the deaths of more than 110 children, the system was expanded nine days later. It shows that the policy document meant to guarantee human judgment, DoD Directive 3000.09, defines neither "appropriate" nor "judgment," leaving the boundary between human decision and machine autonomy so blurred as to be operationally meaningless.
What remains unproven is harder to nail down. We do not know the full classified scope of Maven's weapons interfaces. We do not know how often operators override algorithmic recommendations, or how often they override correctly. We do not know whether the Pentagon has an internal threshold — a number of targets per hour, a number of seconds per decision — at which "human judgment" becomes impossible, and if so, whether that threshold has already been crossed.
We do know this: a system designed to help analysts sort drone footage has become a system that reportedly processes up to 5,000 targets per day. We know that the human role has shrunk from analysis to supervision to a single approval click. And we know that the institution in charge of this system has twice treated "human control" as subordinate to operational speed: once by ejecting a vendor over safety guarantees, and once by expanding Maven after a lethal failure.
Not every error can be rewound. When a lethal system is designed to move faster than humans can verify, you have de facto prioritized killing speed over everything else, including human life.
Works Cited
"Anthropic-Pentagon Battle Shows How Big Tech Has Reversed Course on AI and War." *The Guardian*, 13 Mar. 2026, www.theguardian.com/technology/2026/mar/13/anthropic-pentagon-artificial-intelligence.
Browne, Ryan. "Pentagon Signs Classified AI Deals with Nvidia, Microsoft, and AWS After Ejecting Anthropic Over Safety Limits." *The Next Web*, 6 May 2026, thenextweb.com/news/pentagon-ai-deals-anthropic-safety-limits.
"Cheap Drones, Expensive Lessons: Ethics, Innovation, and Regulation of Autonomous Weapon Systems." *Henry M. Jackson School of International Studies*, University of Washington, 11 July 2025, jsis.washington.edu/news/cheap-drones-expensive-lessons-ethics-innovation-and-regulation-of-autonomous-weapon-systems/.
Department of Defense. *DoD Directive 3000.09: Autonomy in Weapon Systems*. 25 Jan. 2023, www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf.
---. Memo: *Establishment of an Algorithmic Warfare Cross-Functional Team (Project Maven)*. 26 Apr. 2017, nsarchive.gwu.edu/document/18583-national-security-archive-department-defense.
"The Business of War": Google Employees Protest Work for the Pentagon." *The New York Times*, 4 Apr. 2018, www.nytimes.com/2018/04/04/technology/google-letter-ceo-pentagon-project.html.
Kovach, Steve. "Over 3,000 Google Employees Signed a Letter Demanding Google Cancel Its Pentagon Contract." *Vice*, 4 Apr. 2018, www.vice.com/en/article/google-project-maven-protest-letter-killer-ai/.
"2026 Minab School Attack." *Wikipedia*, 18 July 2026, en.wikipedia.org/wiki/2026_Minab_school_attack.
"The School the Algorithm Forgot." *The AI Files*, 2026, theaifiles.app/stories/maven-minab.
"Palmer Luckey's Anduril Won Project Maven Pentagon AI Contract." *The Intercept*, 9 Mar. 2019, theintercept.com/2019/03/09/anduril-industries-project-maven-palmer-luckey/.
"Project Maven." *Wikipedia*, 18 July 2026, en.wikipedia.org/wiki/Project_Maven.
"Project Maven and the Architecture of AI Warfare: Silicon Valley, the Pentagon, and the Transformation of Global Security." *Foreign Affairs Forum*, 31 Mar. 2026, www.faf.ae/home/2026/3/31/project-maven-and-the-architecture-of-ai-warfare-silicon-valley-the-pentagon-and-the-transformation-of-global-security.
Salah, Mohamed. "PROJECT MAVEN | The Architecture of Algorithmic Warfare." *Medium*, 24 Mar. 2026, medium.com/@m.salah2405/project-maven-the-architecture-of-algorithmic-warfare-3ff147e7b520.
Shanahan, Jack. Quoted in "Project Maven to Deploy Computer Algorithms to War Zone by Year's End." *U.S. Department of Defense News*, 21 July 2017, www.defense.gov/News/News-Stories/Article/Article/1254719/.
Wong, Scott. "How a Pentagon Contract Became an Identity Crisis for Google." *The New York Times*, 30 May 2018, www.nytimes.com/2018/05/30/technology/google-project-maven-pentagon.html.
*Some 2018 defense-industry press reports place Booz Allen Hamilton's Maven contract at $885 million rather than $751.5 million. The discrepancy may reflect separate contract vehicles or reporting on different award phases.