# Project Maven The most consequential AI weapons program in history — and the case study that defines every ethical debate in defense tech. What began as a $15M Pentagon pilot for drone surveillance AI became a $1.3B+ Palantir-run system deployed across all U.S. military branches and NATO, capable of closing kill chains in under 90 seconds. > [!note] Program vs. product > **Project Maven** is the Pentagon *program* (the Algorithmic Warfare Cross-Function Team, 2017). **[[Maven Smart System]]** is the deployed Palantir *product* that operationalizes it — the software troops actually use. This page covers the program's history and ethics; see [[Maven Smart System]] for the product's technical capabilities, contract trajectory, and combat-usage metrics. ## Timeline ### Phase 1: Origins (2017) - **April 2017** — Deputy Secretary of Defense Robert O. Work launches the Algorithmic Warfare Cross-Function Team, codenamed "Project Maven." Motivation: China's advances in military AI. - Original mission: Use computer vision and machine learning to help analysts process the overwhelming volume of full-motion drone surveillance video. The DOD was collecting more footage than human analysts could ever watch. ### Phase 2: Google (2017-2019) - **2017** — Pentagon awards Google a contract to provide AI/ML for Maven. Google's cloud division sees it as a gateway to lucrative military work — internal emails projected revenue scaling from $15M to $250M/year. - **March 2018** — The contract leaks publicly. Internal firestorm begins. - **April 2018** — Over 3,000 Google employees sign a petition: *"We believe that Google should not be in the business of war."* Nearly a dozen employees resign. - **June 1, 2018** — Google announces it will not renew the Maven contract when it expires in March 2019. CEO [[Sundar Pichai]] publishes Google's "AI Principles," including a pledge not to build AI for weapons or surveillance. - **June 2018** — Google publishes AI ethics principles. **This moment creates the modern defense tech market.** The walkout proved that Big Tech employees would revolt against military contracts, which opened the door for startups like [[Anduril]] that were built from day one to embrace defense work. See: [[Google Project Maven Walkout]] ### Phase 3: Palantir Takes Over (2019-2024) - **2019** — [[Palantir]] steps in as Maven's primary contractor after Google's exit. Builds the Maven Smart System (MSS) on Palantir's Gotham/Foundry platform. - **May 2024** — Pentagon awards Palantir a $480M five-year contract for Maven Smart System, expanding it across military services globally. ### Phase 4: Massive Expansion (2025-2026) - **April 2025** — NATO acquires Maven Smart System for Allied Command Operations. Sole-source procurement — no competitive bid. Deployed to SHAPE (Belgium) and JFC Brunssum (Netherlands) within 30 days. - **May 2025** — Pentagon raises Maven contract ceiling to **$1.3B** (from $480M), citing "growing demand." DOD designates Maven as an official **program of record** — guaranteeing long-term funding. - **March 2026** — Maven designated core military programme across all armed forces branches. JFC Norfolk integration begins April-May 2026. 18th Airborne Corps is primary tester. ### Phase 5: The Iran War (2026) - **Feb–Apr 2026** — [[Maven Smart System]] becomes the coordinating brain of Operation Epic Fury, the U.S. air campaign of the [[Iran War 2026]] — the largest real-world deployment of AI-assisted targeting in history. - **May 2026** — Pentagon [[Chief Digital and Artificial Intelligence Office]] chief [[Cameron Stanley]] publicly discloses the scale: MSS "leveraged" to run **13,000 targets in 38 days**, with classified usage up 89% and peak consumption near **20 billion tokens a day**. Troops showed an "insatiable appetite" for the tech. See [[Cameron Stanley - insatiable appetite for AI]]. ## What Maven Actually Does Now The Maven Smart System is far beyond the original drone video analysis. It now: 1. **Fuses multi-source intelligence** — satellite imagery, drone feeds, signals intelligence, human intelligence, open-source data — into a single operational picture. 2. **Identifies and tracks targets** — Using AI to detect patterns, flag anomalies, and predict adversary behavior. 3. **Closes kill chains** — Connects sensor data to weapons systems. Reports indicate kill chain times as short as **86 seconds** from detection to strike authorization. 4. **Supports strategic planning** — NATO uses it for warfighting planning, force positioning, and logistics. 5. **Incorporates LLMs** — The system now uses large language models, including Anthropic's Claude (via the [[Palantir-Anthropic-Amazon Defense AI Deal]]), for intelligence analysis and natural language querying of classified data. ## The Kill Chain Debate This is the central ethical flashpoint in defense tech (see [[Kill Chain]] for the underlying doctrine): ### Speed vs. Oversight - Maven can close a kill chain in ~86 seconds. Critics argue meaningful human oversight at that speed is impossible — the "human in the loop" becomes a rubber stamp. - Proponents argue AI reduces civilian casualties by improving targeting accuracy over purely human analysis. ### Accuracy Problem - As of 2024 testing, Maven correctly identified objects at approximately **60% accuracy** — compared to **84% for human analysts**. This means AI is currently *less accurate* than humans at the identification step, even as it accelerates the overall process. - A model trained on biased or incomplete data can flag wrong vehicles, misclassify civilians, or detect threat patterns that don't exist. ### [[Iran]] Strikes (2026) - Early reporting (March 2026) put Maven-connected systems behind **2,000+ strikes over four days** — the opening phase of the campaign. In May 2026, [[Cameron Stanley]] disclosed the fuller total: [[Maven Smart System]] helped run **13,000 targets across the 38-day** Operation Epic Fury air war ([[Iran War 2026]]). This is the largest real-world deployment of AI-assisted targeting in history. - Airwars documented what may be the **first confirmed civilian killed in an AI-assisted strike**. Separately, the first day of the Iran campaign killed at least 175 people at a school adjoining an IRGC base — reportedly due to out-of-date intelligence, underscoring that acceleration does not fix bad inputs. - Critics note that the Trump administration has been eroding civilian casualty tracking mechanisms, and that this accountability gap is now being "programmed into the software." ### The Anthropic Paradox Maven incorporates Claude (Anthropic's AI), creating a contradiction: Anthropic positions itself as the "safety-first" AI company, yet its model is embedded in a system that closes kill chains in under 90 seconds. This led to significant backlash and calls to blacklist Anthropic from AI safety circles. See: [[Palantir-Anthropic-Amazon Defense AI Deal]], [[Dual-Use AI]]. ## Who Benefits The Maven pipeline maps directly onto the defense tech network tracked in this KB: - **[[Palantir]]** — Prime contractor, $1.3B+ contract - **[[Anthropic]]** — Claude AI integrated for intelligence analysis - **[[Amazon]]** — AWS infrastructure for classified cloud - **[[Scale AI]]** — Data labeling and AI testing (via [[Thunderforge Program]]) - **[[Peter Thiel]]** — Palantir co-founder, [[Founders Fund]] investor in multiple connected companies The irony: Google's employee revolt against Maven didn't stop Maven — it just redirected the money from Google (a public company responsive to employee pressure) to Palantir (a company built to embrace this work) and startups (where employees self-select for defense alignment). ## See Also - [[Maven Smart System]] — the deployed Palantir product - [[Cameron Stanley]] — Pentagon CDAO; former Maven chief - [[Chief Digital and Artificial Intelligence Office]] — the government owner - [[Iran War 2026]] — the largest AI-targeting deployment to date - [[Google Project Maven Walkout]] - [[Palantir]] - [[Autonomous Weapons Debate]] - [[Dual-Use AI]] - [[Palantir-Anthropic-Amazon Defense AI Deal]] - [[Thunderforge Program]] - [[Alex Karp]]