Joe Weisenthal and Tracy Alloway host Paul Scharre, executive vice president at the Center for a New American Security and author of two books on AI warfare. Scharre wrote Army of None Autonomous Weapons and the Future of War and more recently Four Battlegrounds Power in the Age of Artificial Intelligence. He previously worked in the Office of the Secretary of Defense and served as an Army Ranger, making him uniquely qualified to discuss the intersection of AI and military operations.

The conversation explores the recent dispute between Anthropic and the Department of Defense over AI usage policies, occurring just as reports emerged that Anthropic's Claude AI is being used for Iran war planning. The discussion covers the spectrum of autonomous weapons, from current AI-assisted targeting to potential future fully autonomous systems, and examines how commercial AI companies are grappling with military applications of their technology.

Defining Autonomous Weapons: A Spectrum of Control

Autonomous weapons exist on a spectrum similar to self-driving cars, with current systems having automated features but humans still making targeting decisions. 'The distinction really is a weapon that is choosing its own targets on the battlefield, and it's not where we are today' - Paul.

Current AI systems like Project Maven use machine learning for image classification to identify objects in drone feeds - 'here's a building, here's a person, here's a vehicle' - representing mature decade-old technology.

The Pentagon's policy, developed by Scharre around 2011, still governs autonomous weapons use today, emerging from the 'accidental robotics revolution' during Iraq and Afghanistan wars.

Anthropic's AI Powers Iran War Planning Through Maven

Anthropic's Claude AI is being used by US military for Iran war planning through the Maven Smart System built by Palantir, helping analysts process massive data sets from over 6,000 sorties flown against Iran.

The AI helps locate mobile targets like 'senior Iranian commanders, mobile missile launchers and air defense systems and drone launchers' by fusing satellite imagery, geolocation data, and signals intelligence.

Large language models enable analysts to interact with data conversationally, asking AI to 'look for intersections' in intelligence and map targets to available aircraft and munitions for strike planning.

The accidental school strike resulted from 'outdated data provided by the Defense Intelligence Agency' where a former military compound had been converted to a school but databases were never updated.

Pentagon vs Tech Companies: Who Sets the Rules

The Pentagon's January 2024 AI strategy requires contracts allowing military to use AI tools for 'any lawful use,' conflicting with tech companies' safety policies that restrict offensive cyber attacks and other applications.

The $200 million Anthropic contract 'is not a lot of money for these AI companies' compared to commercial applications, giving them leverage to refuse military partnerships.

OpenAI stepped in after Anthropic's withdrawal, creating a 'race to the bottom' dynamic where 'the lab with maybe the least amount of safety concern' wins government contracts.

Unlike traditional defense contractors like Lockheed Martin whose 'entire raison d'être is building technology for the government,' commercial AI companies have much larger civilian markets.

Future Autonomous Systems and Escalation Risks

Future autonomous weapons could include loitering munitions that 'search over wider areas' and attack targets independently, similar to 1980s systems like the Tomahawk anti-ship missile that hunted Soviet ships.

AI systems interacting at machine speed could create 'flash crash' scenarios like in financial markets, but 'there's no referee to call time out in war' to prevent escalation.

The military cannot develop AI in-house because 'there's a fierce competition for talent in the AI space' and private enterprise can 'mobilize massive amounts of capital towards building data centers.'

Even in a robot-dominated future, 'militaries are going to need people relatively forward deployed' for command and control, especially against sophisticated adversaries who can jam communications.

The Human Element in Life-and-Death Decisions

The 1983 incident where Soviet officer Stanislav Petrov prevented nuclear war by trusting his 'funny feeling in his gut' about a faulty early warning system illustrates why humans remain essential in critical decisions.

Petrov knew 'the Russian system was new and a lot of Soviet technology doesn't work that great at first,' demonstrating contextual understanding that AI currently lacks about consequences and stakes.

AI could help reduce civilian casualties by identifying strikes near 'schools or hospitals or critical civilian infrastructure' and recommending smaller munitions or higher approval levels.

The risk is humans becoming 'less engaged in this process' where they 'don't feel as morally responsible,' potentially leading to more suffering despite technological precision.