14 August 2018

Algorithms

'Algorithm-assisted decision-making in the public sector: framing the issues using administrative law rules governing discretionary power' by Marion Oswald in (2018) Philosophical Transactions of the Royal Society A comments
This article considers some of the risks and challenges raised by the use of algorithm- assisted decision-making and predictive tools by the public sector. Alongside, it reviews a number of long-standing English administrative law rules designed to regulate the discretionary power of the state. The principles of administrative law are concerned with human decisions involved in the exercise of state power and discretion, thus offering a promising avenue for the regulation of the growing number of algorithm-assisted decisions within the public sector. This article attempts to re- frame key rules for the new algorithmic environment and argues that ‘old’ law – interpreted for a new context – can help guide lawyers, scientists and public sector practitioners alike when considering the development and deployment of new algorithmic tools. 
Introduction 
In 1735, in this very journal, one Reverend Barrow published a short piece, hardly a page in length, in which he surveyed births, deaths and overall population in the parish of Stoke-Damerell in Devon.  He notes that ‘the Number of Persons who died, is one more than half the Number of Children born; and that about 1 in 54 died’ in a year when the ‘General Fever’ infected almost all the inhabitants. He further points out that one of the persons buried was ‘a Foreigner brought from on board a Dutch Ship’ and two more were drowned from on board a Man of War ‘but that the Ships Companies are not included in the Number of Inhabitants.’ This data, together with ‘Experience and Observations, both of my self and better Judges’ leads him to ‘reckon the Parish of Stoke-Damerell as healthful an Air as any in England.’ Fifty-four years later, we find William Morgan (communicated by a Reverend Richard Price) promoting ‘the method of determining, from the real probabilities of life, the value of a contingent reversion in which three lives are involved in the survivorship.’ 
In an age when prospects in society – and lines of credit - might be dependent on one’s ‘great expectations’ of an inheritance, calculating the probability of achieving that inheritance (known to lawyers as contingency reversion) becomes of great interest. For instance, I might transfer a piece of land on the following basis: to my niece for her lifetime, remainder to my nephew and his heirs, but if my nephew dies in the lifetime of my niece, then the land reverts to me and my heirs; I have a ‘reversionary interest’ in the land. The question for my eighteenth century nephew is how to value the sum that might be payable on the contingency that he will survive his sister. The method and calculations proposed by Morgan are set out at length and in considerable detail so as to enable a reader to test and critique them. To this author’s non-expert eye, two points are striking. First, that the calculations appear to be based on group data i.e. on the number of persons living at the age of my nephew, and at the end of first year, second year, third year and so, from the age of my nephew. Secondly, the article goes onto criticise a rule proposed by a certain ‘Mr Simpson’ and points to its results as deviating ‘so widely from the truth as to be unfit for use’ [my emphasis] in some cases producing ‘absurd’ results. 
A modern reader might be tempted to regard these articles as illustrations of a naïve age or to a context long past, or to highlight the lack of causal evidence for Reverend Barrow’s conclusion about the ‘healthful’ nature of his parish. Yet both articles tackle issues with which we remain concerned today: the healthiness (or otherwise) of a community, the reasons behind it and the life expectancy of an individual when compared to others. Risk forecasting and predictive techniques to aid decision- making have become commonplace in our society, not least within public services such as criminal justice, security, benefit fraud detection, health, child protection and social care. We should be better at it than our eighteenth century clergymen. It has become almost unnecessary to say that we now inhabit an information society. Information technologies driven by the flow of digital data have become pervasive and everyday, often leading to the assumption that access to vast banks of (often individualised) digital data, combined with today’s networked computing power and complex algorithmic tools, will lead automatically to greater knowledge and insight, and so to better predictions. 
Knowledge, however, is not the same as information (as many before me have pointed out): Knowledge, Hassan argues, ‘emerges through the open and experiential and diverse (and often intuitive) working and interpreting of raw data and information.’  Reverend Barrow’s conclusion as to the healthfulness of his parish, for instance, was based, not only on the outcome of analysis of raw data, but on additional ‘experience and observations’ of himself and others. Some criticise such human ‘intrusion’ on the data as casting further doubt on the conclusion. Grove and Meehl, a leading proponent of the use of statistical, algorithmic methods of data analysis over clinical methods, argued that ‘To use the less efficient of two prediction procedures in dealing with such matters is not only unscientific and irrational, it is unethical. To say that the clinical-statistical issue is of little importance is preposterous.’  It is this often-claimed superiority, together with the potential for more consistent application of relevant factors often taken from large datasets, that give algorithmic tools their appeal in many public sector contexts. Although this article is written from a legal perspective, it draws attention to arguments made in the ongoing ‘algorithmic predictions versus purely human judgement’ debate and applies these to the legal principles discussed below. It is particularly concerned with algorithm-assisted decisions, whereby an algorithmic output, prediction or recommendation produced by machine learning technique is incorporated into a decision-making process requiring a human to approve or apply it. ‘Machine learning involves presenting the machine with example inputs of the task that we wish it to accomplish. In this way, humans train the system by providing it with data from which it will be able to learn. The algorithm makes its own decision regarding the operation to be performed to accomplish the task in question.’  Machine learning algorithms are ‘probabilistic...their output is always changing depending on the learning basis they were given, which itself changes in step with their use.’ 
Predictive algorithms and administrative law 
The growth in the use of intensive computational statistics, machine-learning and algorithmic methods by the UK public sector shows no sign of abating. What then should be the role of the human when these tools are planned and then deployed, particularly when the accuracy of an algorithmic prediction is claimed to be at least comparable to the accuracy of a human one? I consider this question by reference to a number of connected English administrative law rules, some of which (such as natural justice) date back to before the origins of this journal. I have done this because this body of law governs the exercise of discretionary powers and duties by state bodies, and thus the humans working within them; discretion must be exercised within boundaries or the public body is acting unlawfully. As Le Sueur explains, ‘The assumption made until comparatively recently is that the decision-maker using the executive power conferred by Parliament is a human being or an institution composed of humans and that there is a human who will be accountable and responsible for the decision.’ 
We see this today in witnesses called to give evidence to Parliamentary Select Committees. The introduction of an algorithm to replace, or even only to assist, the human decision-maker represents a challenge to this assumption and thus to the rule of law, and the power of Parliament to decide upon the legal basis of decision-making by public bodies. I argue below however that English administrative law – in particular the duty to give reasons, the rules around relevant and irrelevant considerations and around fettering discretion – is flexible enough to respond to many of the challenges raised by the use of predictive machine learning algorithms, and can signpost key principles for the deployment of algorithms within public sector settings. These principles, although derived from historic case- law, have already been applied and refined to modern government, to the development of the welfare state, privatisation, the development of executive agencies and so on. 
I then attempt to re-frame each of these rules in order to suggest how they could guide future algorithm-assisted decision-making by public bodies affecting rights, expectations and interests of individuals. In doing so, I do not recommend any particular method of building or interpreting these systems - as to do so would require consideration of many different contexts and informational needs - but to suggest principles to guide those engaged in future development work. I focus attention on the requirements of legitimate decision-making from the perspective of the public sector decision-maker, rather than from the perspective of the subject. Fair decision-making in accordance with administrative law rules by its very nature also protects the interests of the human subject of those decisions. I argue that carefully considering exactly what the algorithm is or is not predicting, and explaining to the decision-maker at the point results are displayed, is key to ensuring this fairness.