It has been three years since I last wrote an update on my PhD. I guess another post is in order.
My PhD plan was formally green-lit in October 2019. I am now over three years into this thing. There are roughly two more years left on the clock. I update my plans on a rolling basis. By my latest estimation, I should be ready to request a date for my defense in May 2023.
Of course, the pandemic forced me to adjust course. I am lucky enough not to be locked into particular methods or cases that are fundamentally incompatible with our current predicament. But still, I had to change up my methods, and reconsider the sequencing of my planned studies.
The conference paper I mentioned in the previous update, using the MX3D bridge to explore smart cities’ logic of control and cityness, was rejected by DIS. I performed a rewrite, but then came to the conclusion it was kind of a false start. These kinds of things are all in the game, of course.
The second paper I wrote uses the Transparent Charging Station to investigate how notions of transparent AI differ between experts and citizens. It was finally accepted late last year and should see publication in AI& Society soon. It is titled Tensions in Transparent Urban AI: Designing A Smart Electric Vehicle Charge Point. This piece went through multiple major revisions and was previously rejected by DIS and CHI.
A third paper, Contestable AI by Design: Towards A Framework, uses a systematic literature review of AI contestability to construct a preliminary design framework, is currently under review at a major philosophy of technology journal. Fingers crossed.
And currently, I am working on my fourth publication, tangentially titled Contestable Camera Cars: A Speculative Design Exploration of Public AI Systems Responsive to Value Change, which will be based on empirical work that uses speculative design as a way to develop guidelines and examples for the aforementioned design framework, and to investigate civil servants’ views on the pathways towards contestable AI systems in public administration.
Once that one is done, I intend to do one more study, probably looking into monitoring and traceability as potential leverage points for contestability, after which I will turn my attention to completing my thesis.
Aside from my research, in 2021 was allowed to develop and teach a master elective centered around my PhD topic, titled AI& Society. In it, students are equipped with technical knowledge of AI, and tools for thinking about AI ethics. They apply these to a design studio project focused on conceptualizing a responsible AI-enabled service that addresses a social issue the city of Amsterdam might conceivably struggle with. Students also write a brief paper reflecting on and critiquing their group design work. You can see me on Vimeo do a brief video introduction for students who are considering the course. I will be running the course again this year starting end of February.
I also mentored a number of brilliant master graduation students: Xueyao Wang (with Jacky Bourgeois as chair) Jooyoung Park, Loes Sloetjes (both with Roy Bendor as chair) and currently Fabian Geiser (with Euiyoung Kim as chair). Working with students is one of the best parts of being in academia.
All of the above would not have been possible without the great support from my supervisory team: Ianus Keller, Neelke Doorn and Gerd Kortuem. I should also give special mention to Thijs Turel at AMS Institute’s Responsible Sensing Lab, where most of my empirical work is situated.
If you want to dig a little deeper into some of this, I recently set up a website for my PhD project over at contestable.ai.
Thijs Turèl of AMS Institute and myself presented a version of the talk below at the Cities for Digital Rights conference on June 19 in Amsterdam during the We Make the City festival. The talk is an attempt to articulate some of the ideas we both have been developing for some time around contestability in smart public infrastructure. As always with this sort of thing, this is intended as a conversation piece so I welcome any thoughts you may have.
The basic message of the talk is that when we start to do automated decision-making in public infrastructure using algorithmic systems, we need to design for the inevitable disagreements that may arise and furthermore, we suggest there is an opportunity to focus on designing for such disagreements in the physical objects that people encounter in urban space as they make use of infrastructure.
We set the scene by showing a number of examples of smart public infrastructure. A cyclist crossing that adapts to weather conditions. If it’s raining cyclists more frequently get a green light. A pedestrian crossing in Tilburg where elderly can use their mobile to get more time to cross. And finally, the case we are involved with ourselves: smart EV charging in the city of Amsterdam, about which more later.
We identify three trends in smart public infrastructure: (1) where previously algorithms were used to inform policy, now they are employed to perform automated decision-making on an individual case basis. This raises the stakes; (2) distributed ownership of these systems as the result of public-private partnerships and other complex collaboration schemes leads to unclear responsibility; and finally (3) the increasing use of machine learning leads to opaque decision-making.
These trends, and algorithmic systems more generally, raise a number of ethical concerns. They include but are not limited to: the use of inductive correlations (for example in the case of machine learning) leads to unjustified results; lack of access to and comprehension of a system’s inner workings produces opacity, which in turn leads to a lack of trust in the systems themselves and the organisations that use them; bias is introduced by a number of factors, including development team prejudices, technical flaws, bad data and unforeseen interactions with other systems; and finally the use of profiling, nudging and personalisation leads to diminished human agency. (We highly recommend the article by Mittelstadt et al. for a comprehensive overview of ethical concerns raised by algorithms.)
So for us, the question that emerges from all this is: How do we organise the supervision of smart public infrastructure in a democratic and lawful way?
There are a number of existing approaches to this question. These include legal and regulatory (e.g. the right to explanation in the GDPR); auditing (e.g. KPMG’s “AI in Control” method, BKZ’s transparantielab); procurement (e.g. open source clauses); insourcing (e.g. GOV.UK) and design and engineering (e.g. our own work on the transparent charging station).
We feel there are two important limitations with these existing approaches. The first is a focus on professionals and the second is a focus on prediction. We’ll discuss each in turn.
First of all, many solutions target a professional class, be it accountants, civil servants, supervisory boards, as well as technologists, designers and so on. But we feel there is a role for the citizen as well, because the supervision of these systems is simply too important to be left to a privileged few. This role would include identifying wrongdoing, and suggesting alternatives.
There is a tension here, which is that from the perspective of the public sector one should only ask citizens for their opinion when you have the intention and the resources to actually act on their suggestions. It can also be a challenge to identify legitimate concerns in the flood of feedback that can sometimes occur. From our point of view though, such concerns should not be used as an excuse to not engage the public. If citizen participation is considered necessary, the focus should be on freeing up resources and setting up structures that make it feasible and effective.
The second limitation is prediction. This is best illustrated with the Collinridge dilemma: in the early phases of new technology, when a technology and its social embedding are still malleable, there is uncertainty about the social effects of that technology. In later phases, social effects may be clear but then often the technology has become so well entrenched in society that it is hard to overcome negative social effects. (This summary is taken from an excellent van de Poel article on the ethics of experimental technology.)
Many solutions disregard the Collingridge dilemma and try to predict and prevent adverse effects of new systems at design-time. One example of this approach would be value-sensitive design. Our focus in stead is on use-time. Considering the fact that smart public infrastructure tends to be developed on an ongoing basis, the question becomes how to make citizens a partner in this process. And even more specifically we are interested in how this can be made part of the design of the “touchpoints” people actually encounter in the streets, as well as their backstage processes.
Why do we focus on these physical objects? Because this is where people actually meet the infrastructural systems, of which large parts recede from view. These are the places where they become aware of their presence. They are the proverbial tip of the iceberg.
The use of automated decision-making in infrastructure reduces people’s agency. For this reason, resources for agency need to be designed back into these systems. Frequently the answer to this question is premised on a transparency ideal. This may be a prerequisite for agency, but it is not sufficient. Transparency may help you become aware of what is going on, but it will not necessarily help you to act on that knowledge. This is why we propose a shift from transparency to contestability. (We can highly recommend Ananny and Crawford’s article for more on why transparency is insufficient.)
To clarify what we mean by contestability, consider the following three examples: When you see the lights on your router blink in the middle of the night when no-one in your household is using the internet you can act on this knowledge by yanking out the device’s power cord. You may never use the emergency brake in a train but its presence does give you a sense of control. And finally, the cash register receipt provides you with a view into both the procedure and the outcome of the supermarket checkout procedure and it offers a resource with which you can dispute them if something appears to be wrong.
None of these examples is a perfect illustration of contestability but they hint at something more than transparency, or perhaps even something wholly separate from it. We’ve been investigating what their equivalents would be in the context of smart public infrastructure.
To illustrate this point further let us come back to the smart EV charging project we mentioned earlier. In Amsterdam, public EV charging stations are becoming “smart” which in this case means they automatically adapt the speed of charging to a number of factors. These include grid capacity, and the availability of solar energy. Additional factors can be added in future, one of which under consideration is to give priority to shared cars over privately owned cars. We are involved with an ongoing effort to consider how such charging stations can be redesigned so that people understand what’s going on behind the scenes and can act on this understanding. The motivation for this is that if not designed carefully, the opacity of smart EV charging infrastructure may be detrimental to social acceptance of the technology. (A first outcome of these efforts is the Transparent Charging Station designed by The Incredible Machine. A follow-up project is ongoing.)
We have identified a number of different ways in which people may object to smart EV charging. They are listed in the table below. These types of objections can lead us to feature requirements for making the system contestable.
Because the list is preliminary, we asked the audience if they could imagine additional objections, if those examples represented new categories, and if they would require additional features for people to be able to act on them. One particularly interesting suggestion that emerged was to give local communities control over the policies enacted by the charge points in their vicinity. That’s something to further consider the implications of.
And that’s where we left it. So to summarise:
Algorithmic systems are becoming part of public infrastructure.
Smart public infrastructure raises new ethical concerns.
Many solutions to ethical concerns are premised on a transparency ideal, but do not address the issue of diminished agency.
There are different categories of objections people may have to an algorithmic system’s workings.
Making a system contestable means creating resources for people to object, opening up a space for the exploration of meaningful alternatives to its current implementation.