Fairness, Accountability and Transparency: Decision-Making Algorithms and the Provision of Human Services

Background/Purpose: Algorithms are computer models used to generate predictable output from data that humans cannot analyze alone. Algorithms can take billions of inputs and derive easily interpretable outputs, such as risk scores or group clusters. Increasingly, algorithms are being used to make decisions that have near irreversible effects on people’s lives, such as criminal defendant probation, credit worthiness and housing allocation. Importantly, recent scholarship notes that some of these algorithms are no more accurate than humans with no domain expertise. Human service organizations are also employing algorithms to make decisions: for example, child welfare jurisdictions around the world are implementing algorithms to predict likely outcomes and determine whether to investigate a report or allocate services.

Frameworks to develop, evaluate, and implement algorithmic decision-making have not been discussed in the social work field, with profound implications for the ability of researchers, practitioners, and service users to evaluate the ways in which algorithmic service decisions are being made. This paper offers a conceptual framework to develop, evaluate, and implement algorithmic decision-making in human service provision.

Methods: Using the Fairness, Accountability and Transparency in Machine Learning (FAT/ML) principles for accountable algorithms, the National Association of Social Workers (NASW) Code of Ethics Ethical Standards, as well as social science principles for ethical science, this paper develops a conceptual framework for the ethical development, implementation, and evaluation of algorithmic decision-making in human services. The study then applies the conceptual framework to a test case: the use of algorithmic decision-making (also known as predictive analytics) in child welfare. The framework is applied at each point of the child welfare service continuum, with particular emphasis on investigation and out of home placement.

Results: The conceptual model combines the FAT/ML principles for accountable algorithms with the NASW Code of Ethics standards to offer a five-point approach for evaluating the use of algorithmic-decision-making in human services. The framework applies a common sense approach to algorithmic-decision-making that supports fairness, accountability, and transparency in model development, training, and implementation. The central point of the framework is ensuring accountability: algorithms used in human services must have a central contact person (or persons) responsible for fielding inquiries concerning model development, use, and revision. This “algorithmic spokesperson” should be able to communicate the workings of the algorithm in lay language and should be familiar with alternative decision making processes, in order to showcase how the algorithms fares compared to humans.

Conclusions/Implications: The social work profession is principally concerned with providing social, emotional, financial, political and physical supports to individual, communities and families for whom mainstream avenues of acquiring those supports are non-existent. This concern with the margins requires constant re-evaluation of service provision decision-making systems, to ensure benevolence, non-harming, and respect. This paper offers a first step conceptual framework for ensuring such values can be applied to algorithmic-decision-making in human services.