A new way to test whether an AI system actually controls anything or just seems to

Researchers built a method to measure whether a system genuinely makes decisions that change outcomes, rather than just appearing to do so. It separates what looks like agency from what actually is agency—without needing to argue about consciousness, goals, or intent.

For years, researchers have conflated two different things: whether something persists as an object, and whether it actually controls what happens next. This conflation makes it easy to claim a system is an agent when it's just following rules, and hard to prove it actually makes a difference. The practical effect is clarity. You can now run four concrete, testable checks on any finite system to see if it genuinely steers outcomes or just simulates steering. This matters most for AI systems and autonomous machines, where the difference between 'appears to decide' and 'actually decides' determines how much you can trust it.