Does AI help humans make better decisions? A statistical evaluation framework for experimental and observational...

Area of Science:

• Decision science
• Artificial intelligence
• Legal technology

Background:

• Data-driven algorithms, including artificial intelligence (AI), are increasingly used in society.
• Humans often retain final decision-making authority, especially in high-stakes situations.
• The effectiveness of AI in augmenting human decision-making remains a critical question.

Purpose of the Study:

• To introduce a methodological framework for empirically evaluating AI's impact on human decision-making.
• To compare the performance of human-alone, human-with-AI, and AI-alone decision systems.
• To determine if AI recommendations improve decision accuracy and when to follow them.

Main Methods:

• Developed a framework using standard classification metrics and potential outcomes.
• Employed a single-blinded, unconfounded treatment assignment for AI recommendations.
• Applied the methodology to a randomized controlled trial involving judges and a pretrial risk assessment instrument.

Main Results:

• AI-driven risk assessment recommendations did not enhance judges' accuracy in deciding on cash bail.
• AI-alone systems, including risk assessment scores and large language models, exhibited lower classification performance than human judges.
• The study provides insights into the optimal use of AI in human decision processes.

Conclusions:

• AI-generated recommendations did not improve judicial accuracy in pretrial risk assessment.
• Replacing human judges with AI systems, including large language models, resulted in diminished classification performance.
• The findings suggest caution in deploying AI for high-stakes legal decisions without rigorous empirical validation.