AI-Human confidence alignment slashes decision learning complexity
New paper proves alignment can cut regret from Ω(√|H|·|B|·T) to O(√T log T)
A new paper from Nina Corvelo Benz, Eleni Straitouri, and Manuel Gomez-Rodriguez (UPF Barcelona) tackles a core challenge in AI-assisted decision-making: how should humans learn to trust AI predictions when both sides communicate confidence? The authors formalize the problem as a two-armed online contextual learning problem with full feedback, proving a lower bound of Ω(√(|H|·|B|·T)) on expected regret — where H and B are the sets of human and AI confidence values, and T is time. This establishes the inherent difficulty of learning optimal decisions without structure.
Crucially, when AI and human confidence are perfectly aligned, the regret drops to O(√(|H|·T log T)), and under further conditions to O(√(T log T)). The improvement is exponential in the size of the confidence sets. Experiments on real human-subject data (simple decision tasks with AI assistance) show the results are robust to violations of perfect alignment. The work directly impacts high-stakes domains like medical diagnosis, loan approval, and autonomous driving, where teaching humans to wisely use AI confidence is both critical and costly.
- Lower bound of Ω(√|H|·|B|·T) regret for any learner in AI-assisted binary decisions
- Perfect alignment between AI and human confidence reduces regret to O(√T log T)
- Real human-subject experiments confirm robustness even when alignment is imperfect
Why It Matters
Quantifies how aligning AI confidence with human intuition can dramatically reduce the cost of learning to collaborate.