Human Decision-Making with AI Assistance under Correlated Features

Researchers prove that when AI assists human decision-making with correlated features, stationary recommendation policies perform arbitrarily poorly, requiring instead an explore-then-commit strategy where AI initially recommends diverse options for human learning before committing to optimal selections. The study provides computational complexity results and algorithms for finding near-optimal policies, with exploration duration dependent on feature correlation strength.

This theoretical computer science research addresses a fundamental challenge in human-AI collaboration: how AI systems should guide human learning when features contain dependencies rather than operating independently. The work challenges conventional wisdom by demonstrating that repeatedly recommending the same diagnostic tests or decisions fails catastrophically when those recommendations correlate with each other, forcing humans to learn biased or incomplete feature relationships.

The explore-then-commit framework represents a paradigm shift in AI-assisted decision systems. Rather than optimizing immediate accuracy, the approach deliberately introduces controlled exploration early on, allowing humans to develop accurate mental models of underlying relationships. This balances short-term performance against long-term human capability development—a tension often ignored in AI deployment scenarios. The degree of feature correlation directly determines how long the exploration phase should persist, creating a quantifiable tradeoff between learning and efficiency.

For practical applications spanning healthcare, finance, and enterprise systems, this research validates a suspicion practitioners have observed: blindly following AI recommendations without understanding their reasoning produces brittle, poorly-generalizing human decision-makers. The computational complexity results showing NP-hardness establish fundamental limits on computing optimal policies, while the proposed dynamic programming and approximation algorithms provide implementable solutions for real-world systems with limited planning horizons.

The implications extend beyond academic theory. Organizations deploying AI-assisted decision systems should evaluate feature correlations in their domains and design recommendation strategies that prioritize human learning during initial phases. Future work likely explores how these principles scale to high-dimensional settings and how to measure human learning progress in practice.

• →Stationary AI recommendation policies fail arbitrarily poorly when features are correlated, contradicting assumptions from independent feature settings
• →Optimal policies must follow explore-then-commit structures, with initial diverse recommendations enabling human learning before commitment to single strategies
• →Exploration phase duration depends quantitatively on feature correlation degree, creating measurable tradeoffs between immediate accuracy and long-term capability
• →Computing optimal policies is NP-hard, but dynamic programming and approximation algorithms provide practical near-optimal solutions for finite horizons
• →Feature correlations in real-world domains require deliberate system design to balance short-term decision quality against human learning objectives