AIWill We Know If AI Takes Over? Q&A with Benjamin Boudreaux
“Over time, we would reach a threshold where the options that are presented to us have all been shaped by AI,” says RAND’s Benjamin Boudreaux. “It wouldn’t happen all at once, and it might not even feel like a loss. It could feel like a convenience, a relief. Our decisions might seem to be getting easier,” he says. “And then, really, humans would no longer be in charge.”
Benjamin Boudreaux could feel something slipping in this age of AI. What happens, he wondered, if we off-load our ability to make decisions, to shape the future, to machines? How will we know if we start to lose part of what makes us human, our agency?
Boudreaux is a policy researcher at RAND. His research focuses on navigating the transition to transformative AI. He teamed up with RAND mathematician Alvin Moon to model how AI could erode human agency over time—and what humans can do to stay in control.
Agency “is the capacity that lets us set and pursue every other value that we care about,” Boudreaux said. “And the danger is that there might not be some clear, dramatic moment when we lose it.”
What would losing human agency to AI actually look like?
It wouldn’t happen all at once, and it might not even feel like a loss. It could feel like a convenience, a relief. Our decisions might seem to be getting easier. But over time, we would reach a threshold where the options that are presented to us have all been shaped by AI. And then, really, humans would no longer be in charge.
How do you see this happening?
We identified three mechanisms by which AI could erode human agency. The first is disenfranchisement, where humans are simply automated out of decisions. The second is AI enfranchisement, where AIs effectively become members of the deciding groups.
But the third mechanism is really the most subtle. We call it agenda control. The AI doesn’t cast a vote itself, but it shapes what options the humans see. It would be practically invisible. The options the AI excludes are just not there anymore, so humans wouldn’t even know about them.
What are the warning signs that we might be heading toward that kind of tipping point?
What’s most important here is the trajectory, the trends over time. Every individual move to delegate a decision to AI might make perfect sense. But thousands or millions of these very reasonable decisions add up to a situation in which no human is really deciding anything.
So you might look at, how often do humans override an AI recommendation? How much time and space do humans have to make decisions? What is the scope of an AI’s decisionmaking authority—is it handling one task or many? Those are the trends we want to measure.
