Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AI Persuasion, Bayesian Attribution, and Career Concerns of Decision-Makers

Hanzhe Li, Jin Li, Ye Luo, Xiaowei Zhang

TL;DR

The paper develops a Bayesian framework to study AI persuasion when human decision-makers (DMs) and AI disagree due to attention versus comprehension differences. It shows that AI interpretability shapes how the DM attributes disagreement sources, with attention-based disagreements being more persuasive than comprehension-based ones. A key result is the averaging effect: making AI uninterpretable can increase persuasion by pooling attention and comprehension differences, particularly aiding DMs with lower attention skills and mitigating career-concern distortions. Under career concerns, uninterpretable AI can improve aggregate decision accuracy by enabling low-skill DMs to rely on AI without signaling inferiority, and the work discusses transparency policies and private-use alternatives to further influence outcomes. These insights have practical implications for AI design and organizational information design, suggesting strategic opacity and private AI-consultation can enhance adoption and welfare.

Abstract

This paper studies AI persuasion by distinguishing between two reasons for disagreement: attention differences, where the AI detects features the decision-maker missed, and comprehension differences, where the AI and the decision-maker interpret observed features differently. We show that AI is more effective in persuading the decision-maker when the disagreement is due to attention differences rather than comprehension differences. We also show that the AI's interpretability shapes how the decision-maker attributes the sources of disagreement and, in turn, whether they follow the AI's recommendation. Our main result is that making AI uninterpretable can actually enhance persuasion and, in the presence of career concerns, improve decision accuracy.

AI Persuasion, Bayesian Attribution, and Career Concerns of Decision-Makers

TL;DR

The paper develops a Bayesian framework to study AI persuasion when human decision-makers (DMs) and AI disagree due to attention versus comprehension differences. It shows that AI interpretability shapes how the DM attributes disagreement sources, with attention-based disagreements being more persuasive than comprehension-based ones. A key result is the averaging effect: making AI uninterpretable can increase persuasion by pooling attention and comprehension differences, particularly aiding DMs with lower attention skills and mitigating career-concern distortions. Under career concerns, uninterpretable AI can improve aggregate decision accuracy by enabling low-skill DMs to rely on AI without signaling inferiority, and the work discusses transparency policies and private-use alternatives to further influence outcomes. These insights have practical implications for AI design and organizational information design, suggesting strategic opacity and private AI-consultation can enhance adoption and welfare.

Abstract

This paper studies AI persuasion by distinguishing between two reasons for disagreement: attention differences, where the AI detects features the decision-maker missed, and comprehension differences, where the AI and the decision-maker interpret observed features differently. We show that AI is more effective in persuading the decision-maker when the disagreement is due to attention differences rather than comprehension differences. We also show that the AI's interpretability shapes how the decision-maker attributes the sources of disagreement and, in turn, whether they follow the AI's recommendation. Our main result is that making AI uninterpretable can actually enhance persuasion and, in the presence of career concerns, improve decision accuracy.
Paper Structure (45 sections, 17 theorems, 96 equations, 3 figures)

This paper contains 45 sections, 17 theorems, 96 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Model
  3. Latent Fundamentals for Decision: State, Features, and Critical Dimension
  4. Signals for Decisions: Attention and Comprehension
  5. Decision Process
  6. Preliminaries
  7. AI Persuasion
  8. Interpretable AI
  9. Uninterpretable AI
  10. Career Concerns
  11. Setup
  12. Analysis and Results
  13. Transparency Policies
  14. Extensions
  15. Conclusion
  16. ...and 30 more sections

Key Result

Lemma 1

The DM makes a positive initial decision (i.e., $D=1$) if and only if $X^\mathsf{DM}(C^\mathsf{DM})=1$. The AI makes a positive recommendation (i.e., $A=1$) if and only if $X^\mathsf{AI}(C^\mathsf{AI})=1$.

Figures (3)

  • Figure 1: Two sources of disagreement when the DM's initial decision is negative ($D=0$) and the AI recommends a positive decision ($A=1$)
  • Figure 2: The range of DMs persuaded by the interpretable AI
  • Figure 3: The range of DMs persuaded by the uninterpretable AI

Theorems & Definitions (22)

  • Remark 1
  • Lemma 1: Decision Through Critical Dimension
  • Lemma 2: Interpretable AI: Known Attribution
  • Remark 2
  • Corollary 1: Attention Difference Is More Persuasive
  • Proposition 1: Uninterpretability Enhances Persuasion
  • Remark 3
  • Remark 4
  • Lemma 3
  • Proposition 2: Uninterpretability Improves Accuracy
  • ...and 12 more