"On Becoming Human: An African Notion of Justice and Fairness in Machine Learning"

Sabelo Mhlambi

Machine learning and the collection of large amounts of data for the mediation of the human experience presents several ethical challenges. The centralization of data has led to the centralization of power amongst a few companies, countries and continents and this has contributed to the engineering of society through the engineering of data. Questions of responsibility and the right moral action by those involved in the creation of technology remain unresolved. Technology is deployed in an unequal society and perpetuates dominant biases that affect marginalized communities. The typical conversations and notions of fairness and justice in the tech world and western society are ill-equipped to address the challenges of local and global inequality as they seldom include reparative justice. As machines increasingly become more human stronger concepts of fairness and justice, around cooperation and restorative action, are needed. 

"Fairness in Machine Learning: A Measurement Theory Perspective"

Eran Tal

This paper outlines four principles for a theory of fairness in machine learning, building on recent developments in data ethics, measurement theory, and philosophy of science. Barocas et al. (2018) and Hand (2018) have suggested that the fairness of machine-learning algorithms is intimately tied to the design of measurement procedures. However, these authors consider measurement only at the stage of data collection, and do not discuss the measurements that take place during the generation of a model and the extraction of predictions.

“Predictive Policing: Unbiased and Unfair”

Duncan Purves

In recent years, policing in Los Angeles has undergone a paradigm shift. In addition to relying on crime analyst reports, hunches, and intuition, police on patrol are now guided by crime forecasts generated by predictive algorithms. In 2012, the LAPD began a pilot project to test the efficacy of predictive policing, the use of predictive algorithms trained on massive troves of historical crime data, to anticipate the timing and location of criminal activity. Predictive policing has been heralded by police departments as the new frontier of policing, promising crime reduction gains informed by indisputable hard data. Community organizers and some academics do not share this enthusiasm. They have critiqued predictive policing on the grounds that it is unfair, because its crime forecasts depend on biased arrest data generated from years of discriminatory police practices. 

"Interpretability is a Red Herring: Grappling with “Prediction Policy Problems”

Momin Malik

The goal of ‘interpretability’ fails to grapple with the core paradox of machine learning: that we can make effective predictions on the basis of non-causal correlations. If a machine learning model’s correlations are interpretable but non-causal, then we will be systematically misled if we try to use prior knowledge or intuition about how the world works as a way of validating the model’s operation, or if we try to anticipate when the model might break down under changing conditions of the world, or if we seek to ‘fine tune’ parts of the model that we may interpret as effectively unjust while retaining the model’s integrity. Interpretability may be useful for model diagnostics and debugging, but not for ensuring just usage. For just usage, our focus should instead be on whether a situation is one in which correlations are sufficient: a ‘prediction policy problem.’ If we have such a problem, interpretability is not necessary. Conversely, if we do not have such a case, we should not be using machine learning at all, interpretable or not. But determining whether something is indeed a prediction policy problem may be so difficult as to leave little space for the just use of machine learning when it comes to human systems.

"The Relevance and Irrelevance of Philosophy for Algorithmic Decision-Making"

Rueben Binns

Exploring the role that philosophical theories and concepts like justice, fairness, desert, and responsibility, can play in informing the use of data-driven algorithmic decision-making systems in society. Assuming their aim is not just to pursue intriguing questions about the world, but to change it, what can philosophers engaged in machine learning offer those who are implementing, regulating or suffering from the consequences of these systems? I will argue that while philosophical work has much to offer in this context, putting its insights into practice can easily go wrong. 

"What is the Organizational Counterfactual? Categorical versus Algorithmic Prioritization in U.S. Social Policy"

Rebecca Johnson and Simone Zhang

As public organizations increasingly turn to algorithms to help them allocate resources, two research literatures have emerged. First, social science research has taken a critical role of algorithms’ adoption, arguing that bureaucracies’ use of algorithms exacerbates inequality (Fourcade and Healy, 2013, 2016; O’Neil, 2017; Eubanks, 2018; Brayne, 2017). Second, and partially in response, research largely in computer science focuses on how to measure and improve the fairness by which algorithms prioritize (for a summary, see (Chouldechova, 2017)).

We argue that each literature can be strengthened by comparing algorithmic prioritization to an organizational counterfactual : how U.S. social policies typically prioritize. Under U.S. social policy’s traditional regime, categorical prioritization, organizations and institutions group people into distinct and often visible binary categories like “mothers,“ “veterans,” or “children.”

"The Politics of Data: Discrimination or Justice?"

Joshua Simons and Yonadav Shavit

The concept of discrimination is widely thought to capture what is ‘bad’ about cases of bias in machine learning. Computer scientists and lawyers have developed a range of technical and legal approaches to identify, mitigate and seek redress for machine learning models that may be discriminatory. Meanwhile, technology companies have developed tools and processes which aim to show that their models are not unintentionally discriminatory. We argue that discrimination is the wrong concept to capture what is ‘bad’ about certain cases of bias in machine learning. 

We begin by focusing on the process of machine learning. When data sets necessarily encode structures of power and social inequalities, choices involved in the process of machine learning are, we argue, necessarily political. As a result, opening up the black box of machine learning can tell us something about the political concepts we have to capture and address its potential harms. Discrimination, we argue, does not enable us to satisfactorily distinguish between good and bad cases of bias – all machine learning is in some sense an exercise in discrimination.  

"Pre-Trial Algorithmic Risk Assessments: Value Conflicts, Inherent Limitations, and Harm-Mitigation by Design"

Marc Faddoul and Henriette Ruhrmann

Pretrial detention is a major factor driving incarceration in the U.S. criminal justice system, as 21.6% of the U.S. prison population are pretrial detainees1. Many jurisdictions have thus taken action to move toward risk assessment tools to reduce the pretrial detainee population, address judicial bias in detention decisions, or remedy inequities due to monetary release conditions in the cash bail system. Following a brief background on risk assessment tool use in California, we perform a risk assessment of the Public Safety Assessment (PSA), software used in San Francisco and other jurisdictions to assist judges in deciding whether defendants need to be detained before their trial. Informed by pre-existing literature and interviews with a diverse range of stakeholders – including previously incarcerated individuals – we leverage the Handoff Model, a new theoretical framework to analyze value implications of delegating decision making to technology.