Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

When Should Humans Step In? Optimal Human Dispatching in AI-Assisted Decisions

Lezhi Tan, Naomi Sagan, Lihua Lei, Jose Blanchet

Abstract

AI systems increasingly assist human decision making by producing preliminary assessments of complex inputs. However, such AI-generated assessments can often be noisy or systematically biased, raising a central question: how should costly human effort be allocated to correct AI outputs where it matters the most for the final decision? We propose a general decision-theoretic framework for human-AI collaboration in which AI assessments are treated as factor-level signals and human judgments as costly information that can be selectively acquired. We consider cases where the optimal selection problem reduces to maximizing a reward associated with each candidate subset of factors, and turn policy design into reward estimation. We develop estimation procedures under both nonparametric and linear models, covering contextual and non-contextual selection rules. In the linear setting, the optimal rule admits a closed-form expression with a clear interpretation in terms of factor importance and residual variance. We apply our framework to AI-assisted peer review. Our approach substantially outperforms LLM-only predictions and achieves performance comparable to full human review while using only 20-30% of the human information. Across different selection rules, we find that simpler rules derived under linear models can significantly reduce computational cost without harming final prediction performance. Our results highlight both the value of human intervention and the efficiency of principled dispatching.

When Should Humans Step In? Optimal Human Dispatching in AI-Assisted Decisions

Abstract

AI systems increasingly assist human decision making by producing preliminary assessments of complex inputs. However, such AI-generated assessments can often be noisy or systematically biased, raising a central question: how should costly human effort be allocated to correct AI outputs where it matters the most for the final decision? We propose a general decision-theoretic framework for human-AI collaboration in which AI assessments are treated as factor-level signals and human judgments as costly information that can be selectively acquired. We consider cases where the optimal selection problem reduces to maximizing a reward associated with each candidate subset of factors, and turn policy design into reward estimation. We develop estimation procedures under both nonparametric and linear models, covering contextual and non-contextual selection rules. In the linear setting, the optimal rule admits a closed-form expression with a clear interpretation in terms of factor importance and residual variance. We apply our framework to AI-assisted peer review. Our approach substantially outperforms LLM-only predictions and achieves performance comparable to full human review while using only 20-30% of the human information. Across different selection rules, we find that simpler rules derived under linear models can significantly reduce computational cost without harming final prediction performance. Our results highlight both the value of human intervention and the efficiency of principled dispatching.
Paper Structure (38 sections, 3 theorems, 77 equations, 7 figures, 3 tables)

This paper contains 38 sections, 3 theorems, 77 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Our Contribution
  3. Related Work
  4. Information Acquisition and Human--AI Collaboration
  5. AI and LLMs in Peer Review
  6. Problem Setting
  7. AI and human signals
  8. Selective human querying
  9. Objective
  10. Optimal Selection Rules
  11. Optimal Selection rules in non-parametric setting
  12. Estimation of reward under non-parametric setting: two-stage regression
  13. First-stage regression and cross-fitting.
  14. Orthogonal pseudo-outcome.
  15. Second-stage regression for learning $R_{\pi}(\cdot)$.
  16. ...and 23 more sections

Key Result

Theorem 1

Suppose Assumption assump:rich holds. Then the objective in eq:obj is equivalent to where Equivalently, defining the reward the adaptive problem reduces to and the optimal rule satisfies

Figures (7)

  • Figure 1: High-level diagram of our methodology in the AI-assisted peer review setting. First, an LLM generates a set of feature estimates $A$ for a paper. Next, a selection policy determines which subset of aspects to query from a human reviewer. Finally, the selected human inputs, together with the LLM features, are used to fit a prediction model for the target outcome.
  • Figure 2: MAE of estimated score, plotted with respect to number of aspects selected for human query. For each number of selected aspects, the mean metric value over 20 different training and test splits is plotted, with 95% confidence intervals shown as shaded regions. As selection-free baselines, the performance of models trained on $A$ ("LLM"), $H$ ("Human"), and $(A, H)$ ("Human + LLM") are plotted as dashed lines. For each LLM, results with a linear prediction model (ridge regression) are plotted on the left and a nonlinear model (gradient boosting) are plotted on the right. See Table \ref{['tab:average_metric_values']} for more regression models.
  • Figure 3: The two most selected aspects are Claims_Evidence_Rigor and Originality_Novelty.
  • Figure 4: Prompt for collecting LLM reviews
  • Figure 5: Prompt for tagging the human reviews to generate human features $H$.
  • ...and 2 more figures

Theorems & Definitions (4)

  • Theorem 1: Reduction to contextual reward maximization
  • Theorem 2: Orthogonal two-stage rate
  • Theorem 3: Asymptotic normality of $\hat{R}^\text{(lin)}_\pi$
  • proof