Chapter 1: The Data Dilemma

The data sector faces significant challenges, chief among them the need to address bias in its systems. The demand for inclusive algorithmic practices has emerged as a pressing issue in technology today. Public sentiment is increasingly weary of the negative consequences associated with AI-driven applications, leading to rising public relations costs for tech companies striving to restore their reputations. Both the public and data-driven enterprises recognize the urgency of preventing algorithms from causing harm. Despite the immense potential of AI and algorithmic methods, companies are struggling to maintain or rebuild customer trust.

No single organization possesses the ultimate solution. Each firm is actively refining its data processes to uncover disparities, reduce bias, enhance fairness, and comply with existing and anticipated regulations. The quest for algorithmic solutions that can be effectively measured and scaled while controlling personnel expenses is ongoing.

The existing strategies for fostering a more humane technology landscape can be categorized into two main groups: frameworks and mathematical approaches. The term "mathematical approaches" encompasses any algorithmic model or system utilized by data, AI, and tech firms. Many overlook the strict numerical foundations that underpin our digital ecosystem. Various frameworks are emerging within and beyond tech circles. Notable examples include: The Urban Institute’s Principles for Advancing Equitable Data Practices, We All Count’s Data Equity Framework, and the proposed Algorithmic Accountability Act of 2022’s Impact Assessment template.

The Urban Institute’s Principles for Advancing Equitable Data Practices divides the data lifecycle into stages: acquisition, analysis, dissemination, and disposition. At each stage, it advocates for a comprehensive understanding of data, urging collaboration with affected communities and transparency regarding data limitations.

We All Count’s Data Equity Framework outlines a seven-stage data pipeline: funding, motivation, project design, data collection and sourcing, analysis, interpretation, and communication and distribution. Each stage is designed to raise awareness and intentionally address potential equity concerns. Notably, the final four stages mirror those of the Urban Institute’s principles.

The proposed Algorithmic Accountability Act of 2022 introduces an impact assessment template aimed at enhancing annual transparency and accountability for AI processes and systems in large tech firms. Key areas of focus include performance, fairness, explainability, opportunities for recourse, privacy and security, personal safety, efficiency, and cost.

On the other hand, there is a wide array of mathematical approaches striving to embed inclusive practices throughout the data pipeline. Recent examples include IBM AI Fairness 360, FairML, and Fairlearn.

IBM AI Fairness 360 is an extensive Python toolkit designed for large-scale AI projects, aimed at minimizing various biases' detrimental effects. It includes ten algorithms to mitigate disparities in training datasets, classifiers, and predictive models, along with over 70 fairness-aware metrics.

FairML, developed by Julius Adebayo, offers a toolkit for auditing predictive models to combat discrimination in algorithmic systems. However, it is not considered a distinct subfield of fairness in AI, which may lead to some confusion in online searches.

Fairlearn comprises a suite of Python libraries aimed at algorithmically identifying and mitigating known mathematical disparities in established machine learning techniques. It evaluates the effectiveness of its fairness routines using existing metrics to assess the negative impacts of models and conducts comparative analyses based on accuracy.

With this brief overview of six approaches to promoting inclusion in technology, the landscape can feel daunting. Frameworks provide a broad conceptual understanding of prioritizing human factors but often lack practical implementation guidance. Conversely, mathematical approaches focus intensely on practical applications but may overlook their effects on individuals.

A gap exists in connecting these frameworks with mathematical methods. While frameworks emphasize the need for human involvement in tech development, certain aspects can be automated, where mathematical approaches might be advantageous. A comprehensive inventory and mapping of these frameworks to specific algorithms is necessary to create equitable and inclusive practices throughout the data pipeline, allowing for a better understanding of human limitations and the constraints of mathematical approaches.

The first video, "The 3 Year AI Reset: How To Get Ahead While Others Lose Their Jobs (Prepare Now)" by Emad Mostaque, discusses strategies for navigating the evolving AI landscape and preparing for future challenges in the job market.

The second video, "Why Most Data Projects Fail and How to Avoid It" by Jesse Anderson, explores common pitfalls in data projects and offers insights on how to ensure their success.