Why Your AI Should Know When You're Tired

human-AI cooperation, a new player is emerging that promises to revolutionize the way we work with machines. It's called FALCON, and it doesn't just ask if a machine or a human should make a decision. It asks if the human is too tired to make it.

Learning to defer (L2D) is a concept where AI systems decide when to act autonomously or defer to a human expert. Sounds like a dream team, right? But there's been a flaw in the system: the assumption that human performance is static. Newsflash: it's not. Human performance declines with fatigue, and that's a fact anyone who's worked late nights can tell you.

How FALCON Changes the Game

Enter FALCON, which stands for Fatigue-Aware Learning to Defer via Constrained Optimisation. This isn't just another acronym in the sea of tech buzzwords. FALCON explicitly models human workload and performance using fatigue curves grounded in psychology. It's about time someone thought about the human side of the equation, wouldn't you say?

FALCON uses a Constrained Markov Decision Process (CMDP) to include both task features and cumulative human workload. The goal? Optimize accuracy in human-AI cooperation, all while staying within cooperation budgets. It's trained using PPO-Lagrangian methods. But let's not get lost in technical jargon. The key takeaway is this: FALCON isn't just about doing things faster or cheaper. It's about doing them smarter, with a real understanding of human limits.

Real-World Implications

The introduction of FALCON could mean big changes across industries. Imagine a medical AI that knows when a doctor has been on call too long and needs to take over more of the decision-making process. Or in a manufacturing setting, where AI could step in when workers have had a long shift. The productivity gains went somewhere. Not to wages, mind you, but to smarter decision-making.

But there's a catch. Who pays the cost of implementing these systems? And do we trust AI enough to know when we're not at our best? It's a funny thing, this relationship between humans and machines. We create AI to help us, yet we may not be ready to fully let it take the reins when we're most vulnerable.

The Future of Human-AI Collaboration

FALCON's developers also introduced a benchmark called FA-L2D, which tests how well systems can adapt to varying fatigue levels. Their experiments showed FALCON outperforms current L2D methods consistently, even handling zero-shot generalization to unseen experts with different fatigue patterns.

Ask the workers, not the executives, and you'll find that most people are open to AI stepping in when they're too tired. But it's a delicate balance. Automation isn't neutral. It has winners and losers, and understanding when a system should defer is the first step towards fairer, more effective collaborations.

So, what's the takeaway? AI that understands human fatigue may be the next frontier in collaboration. But as always, the jobs numbers tell one story. The paychecks tell another.