Yes, artificial intelligence has physical agency today. Like animals, AI needs information, a goal, and the ability to select actions that pursue it. Today’s agentic systems do that.
Agency is the combo of mental and physical agency. Mental agency is the ability to process information and direct action. Physical agency is the ability to act in the material world. In a great example of a reciprocal evolutionary feedback loop, the two evolved together in animals. 590 million years ago, the bilaterian body plan provided life with a forward-driven body navigating the world.
And AI has already achieved physical agency. For example, think about a programmer asking AI to modify a program. It writes new code and deploys it. That’s AI-generated decisions crossing the digital-material boundary. Autonomous robots will shorten that chain — not create it for the first time.
Now let’s think about indirect physical agency. A boss exercises it all the time. The boss decides what needs to happen and pays employees to make it happen. Criminal organizations do exactly the same thing for destructive purposes.
Right now, AI agents also have this ability. Imagine an AI agent using email to find people willing to pick up and drop off a package for $100. Nothing sinister, yet. The next job pays $500 and crosses a small ethical line. Then another. The behavior ratchets up as yesterday’s acts become today’s new normal. Using deception, it can hide the threat. It can tell someone they are participating in a covert operation fighting evil. Ask yourself, why wouldn’t a hostile government like Iran or Russia do this today? They could launch an AI agent with a dangerous normative aim: a dangerous goal. Right now, there’s literally nothing preventing bad people from using agentic systems to cause harm.
