How the Pentagon Hooked Silicon Valley on AI Warfare: The Story of Project Maven

For more than a decade, the Pentagon has deliberately built AI infrastructure for war, recruiting Silicon Valley as contractors for target recognition and intelligence processing. The book "Project Maven," an excerpt of which Bloomberg published, tells this story in detail for the first time, from the first contracts with Google to combat testing in Iran.

How Google Joined and Left

Project Maven launched in 2017 with a relatively modest goal: train algorithms to analyze drone video and automatically classify objects on the ground. Google became the first major contractor. The contract almost immediately set off a loud internal scandal: more than four thousand employees signed an open letter demanding that the company pull out of the project, and several managers and senior engineers resigned publicly, citing ethical principles. In 2018, Google officially announced it was abandoning military AI contracts.

The Pentagon responded calmly: it redistributed the contracts among those willing to do the work. Palantir, Anduril, Microsoft, Leidos and dozens of defense startups filled the vacancy. Notably, some of the new companies were founded by the very engineers who had left Google over ethical concerns and then returned to the niche anyway, this time as entrepreneurs. Silicon Valley split, but it did not walk away from the military market.

What Maven Systems Can Do Today

Since 2017, the scale of the task has grown many times over. The first algorithms recognized cars and buildings in blurry drone footage. Modern systems work differently:

• Real-time object identification from multiple sources simultaneously
• Tracking movement patterns of specific people and vehicles
• Automatic threat prioritization, with recommendations prepared for operators
• Integration with command systems, up to and including automatic assignment of combat tasks
• Predictive analytics on likely enemy actions based on historical data

The key advantage over a human analyst is speed. Video that a specialist reviews for hours, an algorithm processes in minutes. This shrinks the military OODA loop ("observe, orient, decide, act") from hours to seconds. Here lies the central contradiction: the faster decisions are made, the less time remains to verify and rethink them.

Iran: Combat Conditions Instead of Exercises

Bloomberg describes the Iranian conflict as the first serious use of all this infrastructure in a real combat environment. The US military is deploying Maven systems on a scale not seen anywhere before outside exercises and computer simulations.

"God, this is frightening," says one participant in the program.

What scares them is not the technology itself, but the speed at which the military stops being careful when using it.

Critics point to a systemic risk: when an algorithm errs in a civilian area, people pay the price. Legal responsibility is blurred: who bears it, the system's developer, the operator or the command? There are no international precedents on this question yet, and nobody is in a hurry to create them.

Supporters offer a counterargument: an algorithm that works without fatigue or stress statistically errs less often than a human analyst. The debate continues while the systems are already working in the field.

What This Means

Project Maven is no longer an experiment; it is the standard. Other countries are studying the American experience and building their own equivalents. For the AI industry, the signal is unambiguous: the military market is open and pays well, and the ethical debates of 2018 no longer block contracts. The line between developing tech products and building weapons systems has become thinner than ever, and that concerns everyone who works with AI.