Algorithmic transparency: seeing through the hype
Dr Jenny Andrew, Head of Data Chartered Society of Physiotherapy
Image courtesy of Joshua Sortino, Unsplash
The UK government’s Central Digital and Data Office (CDDO) has just launched an Algorithmic Transparency Standard for the public sector. The idea is to promote accountability in the algorithmic decision-making that is a growing feature of our public services and civic life. It has largely been received with enthusiasm, and when I agreed to write about it, I thought it would be in a similar vein. This is not the piece I planned…
I should say that I trained as a scientist, and I’m conditioned to view accountability as essential to continuous improvement, and to methodological rigour. And I’ve been an active trade unionist all my working life: I’ve done my time in industrial relations, at the sharp end of holding decision-makers to account. We need more accountability, and structures for citizen-engagement in all our institutions, all our services, and in all businesses. However, the material point, and therefore the object of accountability has to be the decision, and its real-world impacts, not an algorithm or any other tool that feeds it.
To see why this framing matters, look no further than the 2020 A-level results – one of the cases that precipitated the development of the standard. When students chanted “F**k the algorithm”, they gave the exams regulator and the Department for Education a scapegoat for a succession of bad decisions. As a result, the infamous algorithm was dropped, and rightly so, but there’s been relatively little scrutiny of the circumstances that surrounded it.
As it happens, I watched the A-level results fiasco unfold, like a slow-motion car-crash, with commentary from a retired senior examiner and experienced moderator: my mother. “They’ll use the predicted grades in the end,” she said, on the day in March that the exams were cancelled, “They won’t be able to adjust the distribution so far beyond its range.”
Looking back now, my mum’s prediction was a neat illustration of the competencies that weave together into good data design:
- Subject knowledge and broad experience of the real-world phenomena in view: in this case the nature of grade predictions, their relationship to exam grades, and the purposes those A-level grades exist to serve
- Practical knowledge of operational process, including the data inputs, the constraints, and the success criteria for exam moderation
- Solid understanding of statistical methods, such as the relative strengths of calibration within the reference range (e.g., nudging a grade boundary up or down a few points) or extrapolation from a systematically biased distribution
It is rare to find those capabilities tied up in a single person. When we design data-centric processes and tools, therefore, we assemble teams and focus groups, and structure projects with touchpoints that ensure coordination. Data management professionals understand data as a whole-lifecycle concern, so we build relevant expertise into every stage of it.
Bad things happen when we arbitrarily decouple data acquisition from storage and use, or when we pass responsibilities from data manager to software developer to data interpreter like a relay baton. Often the risks in algorithmic decision-making can be traced to those handovers: cliff-edges in domain knowledge and disconnects from the holistic view.
The Algorithmic Transparency Standard, as it stands, reinforces a rather narrow, tech-centric perspective. Here’s how I think it can be recast into a more joined-up approach:Centre the decision, not the algorithm
Even the title is a problem. The current hype has rendered ‘algorithm’ a loaded term, often conflated with AI and machine learning. Although the guidance suggests a wider scope for the standard, I doubt, for example, that an exam moderator working on a spreadsheet would consider using it. (If you’ve seen the mischief that can be done with a spreadsheet, there’s no way you would exempt those decisions from due scrutiny!) The standard itself should be rebalanced to give more weight to the people and process elements that contextualise the analytical technology, and more detail to the data that feeds it.Whole lifecycle view
Analytical technology should be viewed in its place within the data lifecycle. We know that its outputs are only as good as the weakest link in the data supply chain. Taking a whole data lifecycle perspective right from the design phase of the analysis helps to forecast and avert choices that may be embarrassing to recount with hindsight. Furthermore, as any research data manager can attest, designing for accountability makes capture of the essential metadata a lot less troublesome than trying to reconstruct it later.Informed participation
Accountability in public sector decision-making cannot be the preserve of ‘tech-native’ people. We need meaningful participation from across domains and interest groups. Not every citizen will follow all the details of the data and technologies that are used in the public sector. We can, however, target networks and civil sector organisations whose advocates may play a more active role. In the trade union movement, for example, we are developing data literacy among our reps and officers, to complement their expertise in employment and industrial relations at the negotiating table, on behalf of fellow workers.Engage data management experts!
To establish any data protocol in a way that sticks takes a combination of authority, useability and motivation. As a profession, data management can enhance all three in the transparency standard. We are custodians of the organisational structures and processes that will need to support and integrate it. Our experience tells us where contributors will struggle with information gathering, and our workflows are the key to making it easier for them. And our data-centric perspective holds the link between technology and its real-world purpose, keeping it relevant to stakeholders and measuring it by its impacts.
Real-world impact is what matters here. Spare us all from yet more data gathering without material purpose! I wonder how the Algorithmic Transparency Standard will perform outside the 'laboratory conditions’ of its creation. Will we look back in time to see that it made a real-world difference to the decisions affecting society and public services. Probably not with its current, limited viewpoint. Not without expert, structural support.
This isn’t the enthusiastic post I planned to write, not because I want the standard to fail, but because I really want it to succeed. I think it needs a critical friend more than it needs another cheerleader, and our profession is uniquely suited to that brief.
So, I’m thinking about how we can enhance what’s good in the Algorithmic Transparency Standard, how I materialise the principle of accountability in my own professional practice, and how I can support my colleagues and trade union community to adopt it into theirs. I would love to hear other DAMA UK members’ ideas on the subject. And I would love public sector bodies, the CDDO included, to talk to us about how they can build this standard into constructive and sustainable citizen-engagement about the services they provide.