How CDOs can manage algorithmic risks
Automated decision-making algorithms can have many uses in government—but they also entail risks such as bias, errors, and fraud. CDOs can help reduce these risks while capitalizing on these tools’ potential.
The rise of advanced data analytics and cognitive technologies has led to an explosion in the use of complex algorithms across a wide range of industries and business functions, as well as in government. Whether deployed to predict potential crime hotspots or detect fraud and abuse in entitlement programs, these continually evolving sets of rules for automated or semi-automated decision-making can give government agencies new ways to achieve goals, accelerate performance, and increase effectiveness.
However, algorithm-based tools—such as machine learning applications of artificial intelligence (AI)—also carry a potential downside. Even as many decisions enabled by algorithms have an increasingly profound impact, growing complexity can turn those algorithms into inscrutable black boxes. Although often enshrouded in an aura of objectivity and infallibility, algorithms can be vulnerable to a wide variety of risks, including accidental or intentional biases, errors, and fraud.
Read the full CDO Playbook
Create a custom PDF
Subscribe to receive public sector content
Chief data officers (CDOs), as the leaders of their organization’s data function, have an important role to play in helping governments harness this new capability while keeping the accompanying risks at bay.
Understanding the risks
Governments increasingly rely on data-driven insights powered by algorithms. Federal, state, and local governments are harnessing AI to solve challenges and expedite processes—ranging from answering citizenship questions through virtual assistants at the Department of Homeland Security to, in other instances, evaluating battlefield wounds with machine learning-based monitors.1 In the coming years, machine learning algorithms will also likely power countless new Internet of Things (IoT) applications in smart cities and smart military bases.
While such change can be considered transformative and impressive, instances of algorithms going wrong have also increased, typically stemming from human biases, technical flaws, usage errors, or security vulnerabilities. For instance:
- Social media algorithms have come under scrutiny for the way they may influence public opinion.2
- During the 2016 Brexit referendum, algorithms received blame for the flash-crash of the British pound by six percent in two minutes.3
- Investigations have found that an algorithm used by criminal justice systems across the United States to predict recidivism rates is biased against certain racial groups.4
Typically, machine learning algorithms are first programmed and then trained using existing sample data. Once training concludes, algorithms can analyze new data, providing outputs based on what they learned during training and potentially any other data they’ve analyzed since. When it comes to algorithmic risks, three stages of that process can be especially vulnerable:
- Data input: Problems can include biases in the data used for training the algorithm (see sidebar “The problem of algorithmic bias”). Other problems can arise from incomplete, outdated, or irrelevant input data; insufficiently large and diverse sample sizes; inappropriate data collection techniques; or a mismatch between training data and actual input.
- Algorithm design: Algorithms can incorporate biased logic, flawed assumptions or judgments, structural inequities, inappropriate modeling techniques, or coding errors.
- Output decisions: Users can interpret algorithmic output incorrectly, apply it inappropriately, or disregard its underlying assumptions.
The problem of algorithmic bias
Governments have used algorithms to make various decisions in criminal justice, human services, health care, and other fields. In theory, this should lead to unbiased and fair decisions. However, algorithms have at times been found to contain inherent biases, often as a result of the data used to train the algorithmic model. For government agencies, the problem of biased input data constitutes one of the biggest risks they face when using machine learning.
While algorithmic bias can involve a number of factors other than race, allegations of racial bias have raised concerns about certain government applications of AI, particularly in the realm of criminal justice. Some court systems across the country have begun using algorithms to perform criminal risk assessments, an evaluation of the future criminal risk potential of criminal defendants. In nine US states, judges use the risk scores produced in these assessments as a factor in criminal sentencing. However, criminal risk scores have raised concerns over potential algorithmic bias and led to calls for greater examination.5
In 2016, ProPublica conducted a statistical analysis of algorithm-based criminal risk assessments in Broward County, Florida. Controlling for defendant criminal history, gender, and age, the researchers concluded that black defendants were 77 percent more likely than others to be labeled at higher risk of committing a violent crime in the future.6 While the company that developed the tool denied the presence of bias, few of the criminal risk assessment tools used across the United States have undergone extensive, independent study and review. 7
The immediate fallout from algorithmic risks can include inappropriate or even illegal decisions. And due to the speed at which algorithms operate, the consequences can quickly get out of hand. The potential long-term implications for government agencies include reputational, operational, technological, policy, and legal risks.
Taking the reins
To effectively manage algorithmic risks, traditional risk management frameworks should be modernized. Government CDOs should develop and adopt new approaches that are built on strong foundations of enterprise risk management and aligned with leading practices and regulatory requirements. Figure 1 depicts such an approach and its specific elements.
Strategy, policy, and governance
Create an algorithmic risk management strategy and governance structure to manage technical and cultural risks. This should include principles, ethics, policies, and standards; roles and responsibilities; control processes and procedures; and appropriate personnel selection and training. Providing transparency and processes to handle inquiries can also help organizations use algorithms responsibly.
From a policy perspective, the idea that automated decisions should be “explainable” to those affected has recently gained prominence, although this is still a technically challenging proposition. In May 2018, the European Union began enforcing laws that require companies to be able to explain how their algorithms operate and reach decisions.8 Meanwhile, in December 2017, the New York City Council passed a law establishing an Automated Decision Systems Task Force to study the city’s use of algorithmic systems and provide recommendations. The body aims to provide guidance on increasing the transparency of algorithms affecting citizens and addressing suspected algorithmic bias.9
Design, development, deployment, and use
Develop processes and approaches aligned with the organization’s algorithmic risk management governance structure to address potential issues in the algorithmic life cycle from data selection, to algorithm design, to integration, to actual live use in production.
This stage offers opportunities to build algorithms in a way that satisfies the growing emphasis on “explainability” mentioned earlier. Researchers have developed a number of techniques to construct algorithmic models in ways in which they can better explain themselves. One method involves creating generative adversarial networks (GANs), which set up a competing relationship between two algorithms within a machine learning model. In such models, one algorithm develops new data and the other assesses it, helping to determine whether the former operates as it should.10
Another technique incorporates more direct relationships between certain variables into the algorithmic model to help avoid the emergence of a black box problem. Adding a monotonic layer to a model—in which changing one variable produces a predictable, quantifiable change in another—can increase clarity into the inner workings of complex algorithms.11
Monitoring and testing
Establish processes for assessing and overseeing algorithm data inputs, workings, and outputs, leveraging state-of-the-art tools as they become available. Seek objective reviews of algorithms by internal and external parties.
Evaluators can not only assess model outcomes and impacts on a large scale, but also probe how specific factors affect a model’s individual outputs. For instance, researchers can examine specific areas of a model, methodically and automatically testing different combinations of inputs—such as by inserting or removing different parts of a phrase in turn—to help identify how various factors in the model affect outputs.12
The Allegheny County approach
Some governments have begun building transparency considerations into their use of algorithms and machine learning. Allegheny County, Pennsylvania provides one such example. In August 2016, the county implemented an algorithm-based tool—the Allegheny Family Screening Tool—to assess risks to children in suspected abuse or endangerment cases.13 The tool conducts a statistical analysis of more than 100 variables in order to assign a risk score of 1 to 20 to each incoming call reporting suspected child mistreatment.14 Call screeners at the Office of Children, Youth, and Families consult the algorithm’s risk assessment to help determine which cases to investigate. Studies suggest that the tool has enabled a double-digit reduction in the percentage of low-risk cases proposed for review as well as a smaller increase in the percentage of high-risk calls marked for investigation.15
Like other risk assessment tools, the Allegheny Family Screening Tool has received criticism for potential inaccuracies or bias stemming from its underlying data and proxies. These concerns underscore the importance of the continued evolution of these tools. Yet the Allegheny County case also exemplifies potential practices to increase transparency. Developed by academics in the fields of social welfare and data analytics, the tool is county-owned and was implemented following an independent ethics review.16 County administrators discuss the tool in public sessions, and call screeners use it only to decide which calls to investigate rather than as a basis for more drastic measures. The county’s steps demonstrate one way that government agencies can help increase accountability around their use of algorithms.
Are you ready to manage algorithmic risks?
A good starting point for implementing an algorithmic risk management framework is to ask important questions about your agency’s preparedness to manage algorithmic risks. For example:
- What is the potential impact should those algorithms function improperly?
- How well does senior management within your organization understand the need to manage algorithmic risks?
- What is the governance structure for overseeing the risks emanating from algorithms?
Adopting effective algorithmic risk management practices is not a journey that government agencies need to take alone. The growing awareness of algorithmic risks among researchers, consumer advocacy groups, lawmakers, regulators, and other stakeholders should contribute to a growing body of knowledge about algorithmic risks and, over time, risk management standards. In the meantime, it’s important for CDOs to evaluate their use of algorithms in high-risk and high-impact situations and implement leading practices to manage those risks intelligently so that their organizations can harness algorithms to enhance public value.
The rapid proliferation of powerful algorithms in many facets of government operations is in full swing and will likely continue unabated for years to come. The use of intelligent algorithms offers a wide range of potential benefits to governments, including improved decision-making, strategic planning, operational efficiency, and even risk management. But in order to realize these benefits, organizations will likely need to recognize and manage the inherent risks associated with the design, implementation, and use of algorithms—risks that could increase unless governments invest thoughtfully in algorithmic risk management capabilities.