Research map / Large language models / AI and society

Learning to Decide with AI Assistance under Human-Alignment

NeurIPS 2026 · Large language models › AI and society · approach: Human-centered trust and agency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that learning to decide with AI assistance under binary predictions is a two-armed full-feedback contextual online learning problem, proves a regret lower bound, and gives algorithms whose regret improves under perfect human-AI confidence alignment, validated on two human-subject datasets.

Problem

How alignment between AI and human confidence affects the complexity (regret) of learning optimal decision policies through repeated interactions in AI-assisted binary decision making.

Approach

Reduces repeated AI-assisted binary decisions to a two-armed full-feedback contextual online learning problem; under perfect alignment the optimal policy is a monotone threshold in AI confidence, enabling a threshold-learning algorithm whose regret drops from depending on |B| to independent of it, and further to sqrt(T log T) via a generalized DKW inequality.

Key results

• Under perfect alignment the algorithm attains regret O(sqrt(|H| T log T)), and O(sqrt(T log T)) when sqrt(|H|)=O(log T) and B countable.
• Lower bound Omega(sqrt(|H||B|T)) holds for any learner under full feedback.
• On real human-subject data, the proposed algorithm learns the optimal policy and achieves lower regret than the vanilla baseline in all groups/tasks except Census, even where perfect alignment fails.

Surprising findings

• Regret guarantees hold under a weaker monotonicity condition, so the algorithm outperforms the baseline even when perfect alignment is violated.
• On the Census task, where monotonicity fails, the algorithm does not learn the optimal policy as well.

Limitations

• Analysis restricted to binary predictions and binary decisions with full feedback.
• Assumes fixed AI model and degree of alignment over time.
• Monotone-in-both-confidences policy structure not exploited.

Experimental setup

Datasets: Human-Alignment dataset (n=703 card game study) Human-AI Interactions dataset (15,063 predictions from 471 participants; tasks Art, Sarcasm, Cities, Census)

Benchmarks: Human-Alignment dataset Human-AI Interactions dataset

Metrics: expected cumulative regret maximum alignment error (MAE) expected alignment error (EAE)

Baselines: vanilla contextual online learning

Builds on: contextual bandits with policy classes (Agarwal et al.) Dvoretzky-Kiefer-Wolfowitz inequality perfect alignment framework (Straitouri et al.)

Closest papers by meaning

• AI Alignment with Changing and Influenceable Reward Functions (ICML 2024 · similarity 0.88)
• Measuring Error Alignment for Decision-Making Systems (AAAI 2025 · similarity 0.88)
• Model Alignment as Prospect Theoretic Optimization (ICML 2024 · similarity 0.88)
• Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences? (NeurIPS 2025 · similarity 0.88)
• An Evolutionary Perspective on AI Alignment (Student Abstract) (AAAI 2025 · similarity 0.87)
• Sail into the Headwind: Alignment via Robust Rewards and Dynamic Labels against Reward Hacking (ICLR 2025 · similarity 0.87)