Recidivism AI models show less arbitrariness than feared

A new research paper published on arXiv explores the issue of predictive arbitrariness in recidivism risk assessment tools. The study focuses on a machine learning system used for over 15 years, aiming to understand how multiple similarly accurate models can lead to inconsistent predictions for individuals. Researchers developed interpretable models that improved performance and reduced group disparities, finding that structural diversity among models does not always result in significant predictive multiplicity. They propose a simple policy of assigning the lowest risk score among comparable models to mitigate arbitrariness. AI

IMPACT Highlights potential issues with AI in high-stakes decision-making and proposes a mitigation strategy.

RANK_REASON Academic paper on a specific AI application and its potential issues. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →