Introduction
Offering a “barously brief” distillation of Marshall McLuhan’s writings, John M. Culkin expanded on one of McLuhan’s five postulates, Art Imitates Life, with the now-famous line, We shape our tools and thereafter they shape us.Footnote 1 This fear of being shaped and controlled by tools, rather than autonomously wielding them, lies at the heart of current concerns with machine learning and artificial intelligence systems (ML/AI systems). Stories recounting the actual or potential bad outcomes of seemingly blind deference and overreliance on ML/AI systems crowd the popular press. Whether it is Facebook’s algorithms allowing Russian operatives to unleash a weapon of mass manipulation, trained on troves of personal data, on electorates in the US and other countries; inequitable algorithmic bail decisions placing people of color behind bars while whites with similar profiles are sent home to await trial; cars in autonomous mode driving their inattentive could-be-drivers to their death; or algorithms assisting Volkswagen in routing around air quality regulations, there is a growing sense that our tools, if left unchecked, will undermine our choices, our values, and our public policies.
If we fail to grapple with the significant challenges posed by ML/AI systems designed to automate tasks or aid decision making, things may get much worse. At risk are potential decreases in human agency and skill,Footnote 2 both over- and under-reliance on decision support systems,Footnote 3 confusion about responsibility,Footnote 4 and diminished accountability.Footnote 5 Relatedly, as technology reconfigures work practices, it also shifts power in ways that may misalign with liability frameworks, diminishing humans’ agency and control but still leaving them to bear the blame for system failures.Footnote 6 Automation bias, power dynamics, belief in the objectivity and infallibility of data, and distrust of professional knowledge and diminished respect for expertise – all coupled with the growing availability of ML/AI systems and services – portend a potential future in which we are ruled by our tools.
Designing a future in which our tools help us reason and act more effectively, efficiently, and in ways aligned with our social values – i.e., creating the tools that help us act responsibly – requires attention to system design and governance models. ML/AI systems that support us, rather than control us, require designs that foster in-the-moment human engagement with the knowledge and actions systems produce, and governance models that support ongoing critical engagement with ML/AI processes and outputs. Expert decision-support systems are a useful case study to consider the system properties that could maintain human engagement and the governance choices that could ensure they emerge.
We begin by describing three new challenges – design by data, opacity to designer, and dynamic and variable features – posed by the use of predictive algorithmic systems in professional, expert domains. Concerns about inscrutable bureaucratic rules and privatization of public policy making (and the specific opacity that technology can bring to either) apply to predictive machine learning systems generally, but we suggest there are distinctive challenges posed by such predictive systems. We then briefly explore transparency and explainability, two policy objectives that current scholarship suggests are antidotes to such challenges. We show how conceptions of transparency and explainability differ along disciplinary lines (e.g., law, computer science, social sciences) and identify limitations of each concept for addressing the challenges posed by algorithmic systems in expert domains.
We then introduce the concept of contestability and explain the particular benefits of contestable ML/AI systems in the professional context over and above transparent or explainable systems. This approach can be valuable for an algorithmic handoff in a highly professionalized domain, such as the use of predictive coding software – a particular e-discovery tool – by lawyers during litigation. Current governance frameworks around the use of predictive coding in the form of professional norms and codified rules and regulations have their limitations. We argue that an approach centered around contestability would better promote attorneys’ continued, active engagement with these algorithmic systems without relying so heavily on retrospective, case-specific, and costly legal remedies.
The Limitations of Existing Approaches to Protecting Values
Technical systems containing algorithms are shaping and displacing human decision making in a variety of fields, such as criminal justice,Footnote 7 medicine,Footnote 8 product recommendations,Footnote 9 and the practice of law.Footnote 10 Such decision-making handoffs have been met with calls for greater transparency and explainability about system-level and algorithmic processes. The delegation of professional decision making to predictive algorithms – models that predict or estimate an output based on a given inputFootnote 11 – creates additional issues with respect to opacity in machine learningFootnote 12 and to more general concerns with bureaucratic inscrutabilityFootnote 13 and privatization of public power.Footnote 14
Three Challenges Facing Algorithmic Systems in Expert Domains
We identify three challenges facing the use of predictive algorithms in expert systems. First, such predictive algorithms are not designed by technologists in the traditional sense. Whereas engineers of traditional expert systems explicitly program in a set of rules, ideally from the domain knowledge of adept individuals, predictive algorithms supplant this expert wisdom by deriving a set of decision rules from data.
Predictive algorithms can be partitioned into two categories: (1) those focused on outcomes that do not rely too heavily on professional judgment (e.g., was an individual readmitted to the hospital within thirty days of their visit?) versus (2) those focused on outcomes that are more tailored toward emulating the decisions made by professionals with specific domain expertise (e.g., does this patient have pneumonia?). Specifically, the first example can be deemed either true or false simply via observation of admit logs, regardless of professional training. The second example, by way of contrast, is distinct from the first in that it requires medical expertise to make such a diagnosis. In the strictest sense, expert systems fall into the second category,Footnote 15 and as such, inferences of such rules via predictive algorithms create unique challenges for the transfer of expertise from both individuals to the algorithm, and from the algorithm to individuals.
The second challenge is one of opacity. In many ways, this issue is induced by the first. While certain classes of predictive algorithms lend themselves to ease of understanding (such as logistic regression and shallow decision trees), other classes of model make it difficult to understand the rules inferred from the data (such as neural networks and ensemble methods). Unlike expert systems, where domain professionals can review and interrogate the internal rules, the opacity of certain algorithms prevents explicit examination of these decision rules, leaving experts to infer the model’s underlying reasoning from input–output relationships.
Last, these algorithms are case-specific and evolving. They will not necessarily make the same decision about two distinct people in the same way at the same point in time, neither will they necessarily make the same decision about the same individual at varying points in time. This plasticity creates challenges for understanding and interrogating a model’s behavior, as input–output behavior can vary from case to case and can vary over time.
Transparency: Perspectives and Limitations
Due to the challenges described above, algorithmic handoffs have been met with calls for greater transparency.Footnote 16 At a fundamental level, transparency refers to some notion of openness or access, with the goal of becoming informed about the system. However, the word “transparency” lends itself to the question: What is being made transparent?
Given the growing role that algorithmically driven systems are poised to play across government and the private sector, we should exercise care in choosing policy objectives for transparency. A trio of federal laws – two adopted in the 1970s due to fears that the federal government was amassing data about citizens – exemplify three policy approaches to transparency relevant to algorithmic systems. Together, the laws aim to ensure citizens “know what their Government is up to,”Footnote 17 that “all federal data banks be fully and accurately reported to the Congress and the American people,”Footnote 18 that individuals have access to information about themselves held in such data banks, and that privacy considerations inform the adoption of new technologies that manage personal information. These approaches can be summarized as relating to (1) scope of a system, (2) the decision rules of a process, and (3) the outputs.
The Privacy Act of 1974,Footnote 19 which requires notices to be published in the Federal Register prior to the creation of a new federal record-keeping system, and section 208 of the E-Government Act of 2002,Footnote 20 which requires the completion of privacy impact assessments, exemplify the scope perspective. These laws provide notice about the existence and purpose of data-collection systems and the technology that supports them. For example, the Privacy Act of 1974 requires public notice that a system is being created and additional information about the system, including its name and location, the categories of individual and record maintained in the system, the use and purpose of records in the system, agency procedures regarding storage, retrieval, and disposal of the records, etc.Footnote 21 The first tenet of the Code of Fair Information Practices, first set out in a 1973 HEW (Health, Education, Welfare) ReportFootnote 22 and represented in the Privacy Act of 1974 and data-protection laws the world over, stipulates in part that “there must be no personal-data record-keeping systems whose very existence is secret.”Footnote 23 With the Privacy Act of 1974, the transparency theory is one of public notice and scope. Returning to our previous question of “what is being made transparent,” in this approach to transparency, it is precisely the existence and scope being made available.
Unlike the scope aspect of transparency, the decision-rules aspect is not concerned with whether or not such a system exists. Rather, this view of transparency refers to tools to extract information about how these systems function. As an example, consider the Freedom of Information Act (FOIA), a law that grants individuals the ability to access information and documents controlled by the federal government.Footnote 24 The transparency theory here is that the public has a vested interest in accessing such information. But instead of disclosing the information upfront, it sets up a mechanism to meet the public’s demand for it. As such, FOIA allows for individuals to gain access to the decisional rules of these systems and processes. Similarly, the privacy impact assessment requirement of the E-Government Act of 2002 provides transparency around agencies’ consideration of new technologies, as well as their ultimate design choices.
Last, several privacy laws allow individuals to examine the inputs and outputs of systems that make decisions about them. Under this perspective, transparency is not the end goal itself. Rather, transparency supports the twin goals of ensuring fair inputs and understanding the rationale for the outputs by way of pertinent information about the inputs and reasoning. The laws all entitle individuals to access information used about them and to correct or amend data. Some of the privacy laws in this area also entitle individuals to receive information about the reasons behind negative outcomes.Footnote 25 For example, under the Equal Credit Opportunity Act, if a candidate’s credit application is rejected, the credit bureau must provide the key reasons for the decision.Footnote 26 Thus, this type of transparency refers to notice of how a particular decision was reached. These forms of transparency are aimed at individual, rather than collective, understanding; they provide, to a limited extent, insight into the data and the reasoning – or functioning – of systems.
Within the computer science literature, transparency is similar to the functional and outputs perspective presented in law. That is, transparency often refers to some notion of openness around either the internals of a model or system, or around the outputs. Typically, less focus is given to disclosing the subjective choices that were invoked during the system design and engineering process or to system inputs.
The social sciences and statistics, however, take a more comprehensive perspective on transparency. Transparency in these disciplines not only captures the ideas from law and computer science, but also means disclosures about how the data was gathered, how it was cleaned and normalized, the methods used in the analysis, the choice of hyperparameters and other thresholds, etc., often in line with the goals of reproducibility.Footnote 27 The sweep of transparency reflects an understanding that these choices contribute to the methodological design and analysis. This more holistic approach to transparency acknowledges the effect that humans have in this process (reflected in decisions about data, as well as behaviors captured in the data), which is particularly pertinent for predictive algorithms.
Current policy debates, and scientific research, center around explainability and interpretability. Transparency is being reframed, particularly in the computer science research agenda, as an instrumental rather than final objective of regulation and system design. The goal is not to lay bare the workings of the machine, but rather to ensure that users understand how the machines are making decisions – whether those decisions be offering predictions to inform human action or acting independently. This reflects both growing recognition of the inability of humans to understand how some algorithms work even with full access to code and data, but also an emphasis on the overall system – rather than solely the algorithm – as the artifact to be known.
Explainability: Perspectives and Limitations
Explainability is an additional design goal for machine-learning systems. Driven in part by growing recognition of the limits of transparency to foster human understanding of algorithmic systems, and in part by pursuit of other goals such as safety and human compatibility, researchers and regulators are shifting their focus to techniques and incentives to produce machine-learning systems that can explain themselves to their human users. Such desires are well-founded in the abstract. For the purposes of decision making or collaboration, explanations can act as an interface between an end-user and the computer system, with the purpose of keeping a human in the loop for safety and discretion. Hence, explanations invite questioning of AI models and systems to understand limits, build trust, and prevent harm. As with transparency, different disciplines have responded to this call to action by operationalizing both explanations and explainability in differing ways.
One notable use of explanations and explainability comes from the social sciences. MillerFootnote 28 performed a comprehensive literature review of over 200 articles from the social sciences and found that explanations are causal, contrastive, selective, and social. What is pertinent from this categorization is how well the paradigms invoked in predictive algorithms (machine learning, artificial intelligence, etc.) fall within social understandings of explanations. Machine learning raises difficulties for all four of Miller’s attributes of explanations.
For concreteness and clarity, imagine we have a predictive algorithm that classifies a patient’s risk for breast cancer as either low risk, medium risk, or high risk. In this scenario, a causal explanation would answer the question: “Why was the patient classified as high risk?” Alternatively, a contrastive explanation would answer questions of the form, “Why was the patient classified as high risk as opposed to low risk or medium risk?” As such, explanations of the causal type require singular scope on the outcome, whereas contrastive explanations examine not only the predicted outcome, but other candidate alternatives as well.
With respect to machine learning, this distinction is important and suggestive. Machine learning is itself a correlation box. As such, the output itself should not be interpreted as causal. However, when individuals ask for causal explanations of predictive algorithms, they are not necessarily assuming that the underlying data mechanism is causal. Rather, the notion of causality is seeking to understand what caused the algorithm to decide that the patient was high risk, not what caused the patient to be high risk in actuality. Thus, causal explanations can be given of a model built on correlation. However, the fact that they can be produced doesn’t mean that causal explanations further meaningful understanding of the system.
Contrastive explanations are a better fit for machine learning. The very paradigm of machine learning – classification models – are built in a contrastive manner. These models are trained to learn to pick the “best” output given a set of inputs – or equivalently stated, the model is taught to discern an answer to a series of input questions based on the fixed set of alternatives available. Combining these insights, it follows that requiring causal explanations for classification models is inappropriate for determining why a model predicted the value it did. Contrastive explanations, which provide insight into the counterfactual alternatives that the model rejected as viable, transfer more knowledge about the system, than causal ones.
Regardless of whether the type of explanation is causal or contrastive, Miller argued that explanations in the social sciences were selective. That is, explanations tend to highlight a few key justifications rather than being completely exhaustive. Consider the case of a doctor performing a breast cancer-screening test in the absence of a predictive algorithm. When relaying the rationale of their diagnosis to a patient, a doctor would provide sufficient reasons for their decision to justify their answer. Now, consider the state of the world where a handoff has been made to the predictive model. Suppose the model being used relies on 500 features. When explaining why the model predicted the outcome it did, it is indeed unreasonable to assume that providing information about all 500 features would practically relay any information about why the model made the choice it did. As such, requiring explanations of predictive models requires honing into the relevant features of a decision problem, which may differ from patient to patient and may vary over time.
On the aspect of explanations being social, Miller noted that explanations are meant to transfer knowledge from one individual to another. In the example above, where the doctor performs the breast cancer-screening test, this was the point of having the doctor justify their diagnosis to the patients – to inform the patient about their breast cancer-risk level. When applied to technical systems, the goal is to transfer knowledge about the internal logic of how the system reached its conclusion to some individual (or class of individuals). In the case of our breast cancer-risk prediction, this would manifest itself as a way to justify why the algorithm predicted high risk as opposed to low risk. It is worth noting that for predictive algorithms, it is often difficult to truly achieve the social goal of explanations. Certain qualities of algorithms – such as their functional form (e.g., nonlinear, containing interaction terms), their input data, and other characteristics – make it particularly difficult to assess the internal logic of the algorithm itself, or for the system to even explain what it is doing. It is therefore difficult for these machine systems to transfer knowledge to individuals in the form of an explanation that is either causal or contrastive. To the extent that explanations are aimed at improving human understanding of the logic of algorithms, the qualities of some algorithms may be incompatible with this means of transferring knowledge. It may be that the knowledge transfer must come the other way around, from the human to the machine, which is then bound to particular way or ways of knowing.Footnote 29
Thus, there are tensions between the paradigms of predictive algorithms and those characteristics laid out by Miller. As such, the discussion above suggests that our target is off. That is, to actually fully and critically engage with predictive algorithms, this suggests that we require something stronger than transparency and explainability. Enter contestability – the ability to challenge machine predictions.
Toward Contestability as a Feature of Expert Decision-Support Systems
Contestability fosters engagement rather than passivity, questioning rather than acquiescence. As such, contestability is a particularly important system quality where the goal is for predictive algorithms to enhance and support human reasoning, such as decision-support systems. Contestability is one way “to enable responsibility in knowing”Footnote 30 as the production of knowledge is spread across humans and machines. Contestability can support critical, generative, and responsible engagement between users and algorithms, users and system designers, and ideally between users and those subject to decisions (when they are not the users), as well as the public.
Efforts to make algorithmic systems knowable respond to the individual need to understand the tools one uses, as well as the social need to ensure that new tools are fit for purpose. Contestability is a design intervention that can contribute to both.Footnote 31 However, our focus here is on its potential contribution to the creation of governance models that “support epistemically responsible behavior”Footnote 32 and support shared reasoning about the appropriateness of algorithmic systems behavior.Footnote 33
Contestability, the ability to contest decisions, is at the heart of legal rights that afford individuals access to personal data and insight into the decision-making processes used to classify them,Footnote 34 and it is one of the interests that transparency serves. Contestability as a design goal, however, is more ambitious and far-reaching. A system designed for contestability would protect the ability to contest a specific outcome, consistent with privacy and consumer protection law. It would also facilitate generative engagement between humans and algorithms throughout the use of the machine-learning system and support the interests and rights of a broader range of stakeholders – users, designers, as well as decision subjects – in shaping its performance.
Hirsch et al. set out contestability as a design objective to address myriad ethical risks posed by the potential reworking of relationships and redistribution of power caused by the introduction of machine-learning systems.Footnote 35 Based on their experience designing a machine-learning system for psychotherapy, Hirsch et al. offer three lower-level design principles to support contestability: (1) improving accuracy through phased and iterative deployment with expert users in environments that encourage feedback; (2) heightening legibility through mechanisms that “unpack aggregate measures” and “trac[e] system predictions all the way down” so that “users can follow, and if necessary, contest the reasoning behind each prediction”; and relatedly, in an effort to identify and vigilantly prevent system misuse and implicit bias, (3) identifying “aggregate effects” that may imperil vulnerable users through mechanisms that allow “users to ask questions and record disagreements with system behavior” and engage the system in self-monitoring.Footnote 36 Together, these design principles can drive active, critical, real-time engagement with the reasoning of machine-learning system inputs, outputs, and models.
This sort of deep engagement and ongoing challenge and recalibration of the reasoning of algorithms is essential to yield the benefits of humans and machines reasoning together. Concerns that engineers will stealthily usurp or undermine the decision-making logics and processes of other domains have been an ongoing and legitimate complaint about decision support and other computer systems.Footnote 37 Encouraging human users to engage and reflect on algorithmic processes can reduce the risk of stealthy displacement of professional and organizational logics by the logics of software developers and their employers. Where an approach based on explanations imagines questioning and challenging as out-of-band activities – exception handling, appeals processes, etc. – contestable systems are designed to foster critical engagement within the system. Such systems use that engagement to iteratively identify and embed domain knowledge and contextual values, as decision making becomes a collaborative effort within a sociotechnical system.
In the context of decision-support systems, increasing system explainability and interpretability is viewed as a strategy to address errors that stem from automation bias and to improve trust.Footnote 38 Researchers have examined the impact of various forms of explanatory material, including confidence scores, and comprehensive and selective lists of important inputs, on the accuracy of decisions, deviation from system recommendations, and trust.Footnote 39 The relationship between explanations and correct decision making is not conclusive.Footnote 40
Policy debates, like the majority of research on interpretable systems, envision explanations as static.Footnote 41 Yet, the responsive and dynamic tailoring at which machine learning and AI systems excel could allow explanations to respond to the expertise and other context-specific needs of the user, yielding decisions that leverage, and iteratively learn from, the situated knowledge and professional expertise of users.
The human engagement contestable systems invite would align well with regulatory and liability rules that seek to keep humans in the loop. For example, the Food and Drug Administration is directed to exclude from the definition of “device” those clinical decision support systems whose software function is intended for the purpose of:
supporting or providing recommendations to a health care professional about prevention, diagnosis, or treatment of a disease or condition; and enabling [providers] to independently review the basis for such recommendations … so that it is not the intent that such [provider] rely primarily on any of such recommendations to make a clinical diagnosis or treatment decision regarding an individual patient.Footnote 42
By excluding systems that prioritize human discretion from onerous medical-device approval processes, Congress shows its preference for human expert reasoning. Similarly, where courts have found professionals exhibiting overreliance on tools, they have structured liability to foster professional engagement and responsibility.Footnote 43 Systems designed for contestability invite engagement rather than delegation of responsibility. They can do so through both the provision of different kinds of information and an interactive design that encourages exploration and querying.
Professionals appropriate technologies differently, employing them in everyday work practice, as informed by routines, habits, norms, values and ideas and obligations of professional identity. Drawing attention to the structures that shape the adoption of technological systems opens up new opportunities for intervention. Appropriate handoffs to, and collaborations with, decision-support systems demand that they reflect professional logics and provide users with the ability to understand, contest, and oversee decision making. Professionals are a potential source of governance for such systems, and policy should seek to exploit and empower them, as they are well-positioned to ensure ongoing attention to values in handoffs and collaborations with machine-learning systems.
Regulatory approaches should seek to put professionals and decision support systems in conversation, not position professionals as passive recipients of system wisdom who must rely on out-of-system mechanisms to challenge them. For these reasons, calls for explainability fall short and should be replaced by regulatory approaches that drive contestable design. This requires attention to both the information demands of professionals – what they need to know such as training data, inputs, decisional rules, etc. – and processes of interaction that elicit professional expertise and allow professionals to learn about and shape machine decision making.
Contestable Design Directions
Contestable design is a research agenda, not a suite of settled techniques to deploy. The question of what information and interactions will prompt appropriate engagement and shaping of a predictive coding system by professionals is likely to be both domain- and context-specific. However, there are systems in use and under development that support real-time questioning, curiosity, and scrutiny of machine learning systems’ reasoning. First, Google’s People and AI Research (PAIR) Initiative’s “What-if Tool” is an actual tool that allows users to explore a machine-learning model. For example, users can see how changes in aspects of a dataset influence the learned model, understand how different models perform on the same dataset, compare counterfactuals, and test particular operational constraints related to fairness.Footnote 44 Second, LIME (Local Interpretable Model-agnostic Explanations), which generates locally interpretable models to explain the outputs of predictive systems, and SP-LIME, which builds on LIME to provide insight into the model (rather than a given prediction) by identifying and explaining a set of representative instances of the model’s performance, offer information that, if presented to users, could inform their interaction with the model.Footnote 45 While the tools themselves focus only on surfacing information about decisions and models, if integrated with an interactive user interface, they could promote the explorations of predictions and models necessary for sound use of predictive systems to inform professional judgement.
Other research is exploring the ways in which structured interaction between domain experts and predictive models can improve performance.Footnote 46 There are two distinct approaches. One approach enables interaction during the development process. Here, the machine-learning training process is reframed as an HCI task, allowing a set of users the ability to iteratively refine a model during its conception.Footnote 47 In contrast to interaction during the development process, the second approach has focused on ways in which subject matter experts, with domain-specific knowledge, can interact with predictive systems that have already been developed in real time to invoke collaboration, exploration of data, and introspection.Footnote 48 At the very least, ensuring that decisions about things such as thresholds are decided by professionals in the context of use (and remain visible to those using the system), rather than set as defaults, can support greater engagement with predictive systems.
Conclusion
Contestability allows professionals, not just data, to train systems. In doing so, contestability transfers knowledge about how the machine is reasoning to the professional, and it allows the professional to collaborate, critique, and correct the predictive algorithm. While relevant professional norms, ethical obligations, and laws are necessary, design has a role to play in promoting responsible introduction of predictive ML/AI systems in professional, expert domains. Such systems must be designed with contestability in mind from the outset. Designing for contestability has some specific advantages compared to rules and laws. Opportunities to reflect on the inputs and assumptions that shape systems can avert disasters where they misalign with the conditions or understandings of professional users. Reminders of professional responsibilities and potential risks of not complying with them can prompt engagement before undesirable outcomes occur. Contestable design can confer training benefits allowing users to learn through use. Finally, it can be used to signal the distribution of responsibility from the start rather than relying solely on litigation to retrospectively mete it out in light of failures. Contestability can foster professional engagement with tools rather than deferential reliance. To the extent the goal is to yield the best of human-machine knowledge production, designing for contestability can promote the responsible production of knowledge with machine learning tools within professional contexts.