[Paper Review] Measuring Human Adaptation to AI in Decision Making: Application to Evaluate Changes after AlphaGo

Minkyu Shin, Jin Kim+1 more|arXiv (Cornell University)|Dec 30, 2020

Explainable Artificial Intelligence (XAI)23 references4 citations

TL;DR

This paper introduces the Human-AI Gap, a measure to quantify human adaptation to AI in decision-making by comparing human move quality to superhuman AI (e.g., AlphaGo) in Go. It finds that observing AI's reasoning processes—not just its moves—drives meaningful learning, while the measure also detects AI-assisted cheating, demonstrating broad applicability in domains like medicine and education.

ABSTRACT

Across a growing number of domains, human experts are expected to learn from and adapt to AI with superior decision making abilities. But how can we quantify such human adaptation to AI? We develop a simple measure of human adaptation to AI and test its usefulness in two case studies. In Study 1, we analyze 1.3 million move decisions made by professional Go players and find that a positive form of adaptation to AI (learning) occurred after the players could observe the reasoning processes of AI, rather than mere actions of AI. These findings based on our measure highlight the importance of explainability for human learning from AI. In Study 2, we test whether our measure is sufficiently sensitive to capture a negative form of adaptation to AI (cheating aided by AI), which occurred in a match between professional Go players. We discuss our measure's applications in domains other than Go, especially in domains in which AI's decision making ability will likely surpass that of human experts.

Motivation & Objective

• To develop an objective, quantifiable measure of how humans adapt to AI in decision-making contexts.
• To investigate whether and when human experts improve decision quality after exposure to superhuman AI, such as AlphaGo.
• To assess whether the proposed measure can detect negative adaptation, such as AI-assisted cheating in competitive settings.
• To evaluate the broader applicability of the measure in domains where AI outperforms humans but humans remain final decision-makers.
• To explore factors influencing differential rates of adaptation to AI across individuals or groups.

Proposed method

• Define the Human-AI Gap as the difference in decision quality between human players and a superhuman AI, using reinforcement learning-based AI outputs as the gold standard.
• Use historical professional Go game data (1.3 million moves) to compare human move quality before and after AI emergence.
• Compare human decisions to AI’s move quality using a learned policy from AlphaGo, treating it as the benchmark for optimal play.
• Analyze changes in the Human-AI Gap over time and across player groups to detect learning or cheating patterns.
• Apply the measure to detect anomalies in move quality that suggest AI assistance, such as sudden performance spikes.
• Use statistical modeling to assess whether improvements in human decision quality correlate with access to AI reasoning, not just actions.

Experimental results

Research questions

• RQ1 Does exposure to AI reasoning processes lead to measurable improvement in human decision-making quality compared to exposure to only AI actions?
• RQ2 When and where in a game does human adaptation to AI most prominently occur?
• RQ3 Can the Human-AI Gap measure detect instances of AI-assisted cheating in professional Go matches?
• RQ4 How does access to AI reasoning affect adaptation rates among different groups of human experts?
• RQ5 To what extent can the Human-AI Gap be generalized to other domains such as medicine, education, or business?

Key findings

• Human experts significantly improved decision quality after gaining access to AI’s reasoning processes, not merely its move outputs.
• The most substantial learning occurred in the early to middle stages of Go games (moves 1–50), where human players began emulating AI strategies.
• Experts with access to AI reasoning showed a measurable reduction in the Human-AI Gap, indicating faster adaptation compared to those without access.
• The Human-AI Gap measure successfully detected a known case of AI-assisted cheating, where a player’s move quality suddenly aligned with AI output.
• The measure revealed that explainability—understanding *why* AI made a move—was critical for effective human learning, not just observing the outcome.
• The Human-AI Gap is sensitive enough to detect both positive adaptation (learning) and negative adaptation (cheating), supporting its use in diverse, high-stakes decision contexts.