On how to think about large language models

How should we think about large language models (LLMs)? People commonly think and talk about them in terms of human intelligence. To the extent this metaphor does not accurately reflect the properties of the technology, this may lead to misguided diagnoses and prescriptions. It seems to me an LLM is not like a human or a human brain in so many ways. One crucial distinction for me is that LLMs lack individuality and subjectivity.

What are organisms that similarly lack these qualities? Coral polyps and Portuguese man o’ war come to mind, or slime mold colonies. Or maybe a single bacterium, like an E. coli. Each is essentially identical to its clones, responds automatically to chemical gradients (bringing to mind how LLMs respond to prompts), and doesn’t accumulate unique experiences in any meaningful way.

Considering all these examples, the meme about LLMs being like a shoggoth (an amorphous blob-like monster originating from the speculative fiction of Howard Philips Lovecraft) is surprisingly accurate. The thing about these metaphors though is that it’s about as hard to reason about such organisms as it is to reason about LLMs. So to use them as a metaphor for thinking about LLMs won’t work. A shoggoth is even less helpful because the reference will only be familiar to those who know their H.P. Lovecraft.

So perhaps we should abandon metaphorical thinking and think historically instead. LLMs are a new language technology. As with previous technologies, such as the printing press, when they are introduced, our relationship to language changes. How does this change occur?

I think the change is dialectical. First, we have a relationship to language that we recognize as our own. Then, a new technology destabilizes this relationship, alienating us from the language practice. We no longer see our own hand in it. And we experience a lack of control over language practice. Finally, we reappropriate this language use in our practices. In this process of reappropriation, language practice as a whole is transformed. And the cycle begins again.

For an example of this dialectical transformation of language practice under the influence of new technology, we can take Eisenstein’s classic account of the history of the printing press (1980). Following its introduction many things changed about how we relate to language. Our engagement with language shifted from a primarily oral one to a visual and deliberative one. Libraries became more abundantly stocked, leading to the practice of categorization and classification of works. Preservation and analysis of stable texts became a possibility. The solitary reading experience gained prominence, producing a more private and personal relationship between readers and texts. Concerns about information overload first reared its head.

All of these things were once new and alien to humans. Now we consider them part of the natural order of things. They weren’t predetermined by the technology, they emerged through this active tug of war between groups in society about what the technology would be used for, mediated by the affordances of the technology itself.

In concrete material terms, what does an LLM consist of? An LLM is just numerical values stored in computer memory. It is a neural network architecture consisting of billions of parameters in weights and biases, organized in matrices. The storage is distributed across multiple devices. System software loads these parameters and enables the calculation of inferences. This all runs in physical data centers housing computing infrastructure, power, cooling, and networking infrastructure. Whenever people start talking about LLMs having agency or being able to reason, I remind myself of these basic facts.

A printing press, although a cleverly designed, engineered, and manufactured device, is similarly banal when you break it down to its essential components. Still, the ultimate changes to how we relate to language have been profound. From these first few years of living with LLMs, I think it is not unreasonable to think they will cause similar upheavals. What is important for me is to recognize how we become alienated from language, and to see ourselves as having agency in reappropriating LLM-mediated language practice as our own.

On mapping AI value chains

This post was taken down on 26 October 2026. A revised version will appear at this address.

About the original

Report on a workshop held at CSCW 2024 (November 2024). The original ran to about 761 words and 1 image.

Abstract

Generated by Claude, a language model, from the original text; not written by the author.

The workshop, “From Stem to Stern: Contestability Along AI Value Chains”, looked at contesting AI not only at the point of individual decisions but across the whole chain, from resource extraction to deployment. The post shares the AI value chain mapping canvas (CC-BY, Yurrita & Alfrink) used in the first group activity, explains how the 50-minute mapping exercise with ChatGPT as the example case was run, and names its influences (value chain mapping, Wardley mapping). Takeaways include the vocabulary gap between legal and technical communities, the need to move from individual grievance to collective and upstream contestation, and the importance of Global South perspectives. It credits the full workshop team.

On autonomy, design, and AI

In my thesis, I use autonomy to build the normative case for contestability. It so happens that this year’s theme at the Delft Design for Values Institute is also autonomy. On October 15, 2024, I participated in a panel discussion on autonomy to kick things off. I collected some notes on autonomy that go beyond the conceptualization I used in my thesis. I thought it might be helpful and interesting to collect some of them here in adapted form.

The notes I brought included, first of all, a summary of the ecumenical conceptualization of autonomy concerning automated decision-making systems offered by Alan Rubel, Clinton Castro, and Adam Pham (2021). They conceive of autonomy as effective self-governance. To be autonomous, we need authentic beliefs about our circumstances and the agency to act on our plans. Regarding algorithmic systems, they offer this notion of a reasonable endorsement test—the degree to which a system can be said to respect autonomy depends on its reliability, the stakes of its outputs, the degree to which subjects can be held responsible for inputs, and the distribution of burdens across groups.

Second, I collected some notes from several pieces by James Muldoon, which get into notions of freedom and autonomy that were developed in socialist republican thought by the likes of Luxemburg, Kautsky, and Castoriadis (2020, 2021a, 2021b). This story of autonomy is sociopolitical rather than moral. This approach is quite appealing for someone interested in non-ideal theory in a realist mode like myself. The account of autonomy Muldoon offers is one where individual autonomy hinges on greater group autonomy and stronger bonds of association between those producing and consuming technologies. Freedom is conceived of as collective self-determination.

And then third and finally, there’s this connected idea of relational autonomy, which to a degree is part of the account offered by Rubel et al., but in the conceptions here more radical in how it seeks to create distance from liberal individualism (e.g., Christman, 2004; Mhlambi & Tiribelli, 2023; Westlund, 2009). In this, individual capacity for autonomous choice is shaped by social structures. So freedom becomes realized through networks of care, responsibility, and interdependence.

That’s what I am interested in: accounts of autonomy that are not premised on liberal individualism and that give us some alternative handle on the problem of the social control of technology in general and of AI in particular.

From my point of view, the implications of all this for design and AI include the following.

First, to make a fairly obvious but often overlooked point, the degree to which a given system impacts people’s autonomy depends on various factors. It makes little sense to make blanket statements about AI destroying our autonomy and so on.

Second, in value-sensitive design terms, you can think about autonomy as a value to be balanced against others—in the case where you take the position that all values can be considered equally important, at least in principle. Or you can consider autonomy more like a precondition for people to live with technology in concordance with their values, making autonomy take precedence over other values. The sociopolitical and relational accounts above point in this direction.

Third, suppose you buy into the radical democratic idea of technology and autonomy. In that case, it follows that it makes little sense to admonish individual designers about respecting others’ autonomy. They may be asked to privilege technologies in their designs that afford individual and group autonomy. But designers also need organization and emancipation more often than not. So it’s about building power. The power of workers inside the organizations that develop technologies and the power of communities that “consume” those same technologies. 

With AI, the fact is that, in reality, in the cases I look at, the communities that AI is brought to bear on have little say in the matter. The buyers and deployers of AI could and should be made more accountable to the people subjected to AI.

Towards a realist AI design practice?

This is a version of the opening statement I contributed to the panel “Evolving Perspectives on AI and Design” at the Design & AI symposium that was part of Dutch Design Week 2024. I had the pleasure of joining Iohanna Nicenboim and Jesse Benjamin on stage to explore what could be called the post-GenAI possibility space for design. Thanks also to Mathias Funk for moderating.

The slide I displayed:

My statement:

  1. There’s a lot of magical thinking in the AI field today. It assumes intelligence is latent in the structure of the internet. Metaphors like AGI and superintelligence are magical in nature. AI practice is also very secretive. It relies on demonstrations. This leads to a lack of rigor and political accountability (cf. Gilbert & Lambert in VentureBeat, 2023).
  2. Design in its idealist mode is easily fooled by such magic. For example, in a recent report, the Dutch Court of Audit states that 35% of government AI systems are not known to meet expectations (cf. Raji et al., 2022).
  3. What is needed is design in a realist mode. Realism focuses on who does what to whom in whose interest (cf. Geuss, 2008, 23 in von Busch & Palmås, 2023). Applied to AI the question becomes who gets to do AI to whom? This isn’t to say we should consider AI technologies completely inert. It mediates our being in the world (Verbeek, 2021). But we should also not consider it an independent force that’s just dragging us along.
  4. The challenge is to steer a path between, on the one hand, wholesale cynical rejection and naive, optimistic, unconditional embrace, on the other hand.
  5. In my own work, what that looks like is to use design to make things that allow me to go into situations where people are building and using AI systems. And to use those things as instruments to ask questions related to human autonomy, social control, and collective freedom in the face of AI.
  6. The example shown is an animated short depicting a design fiction scenario involving intelligent camera cars used for policy execution in urban public space. I used this video to talk to civil servants about the challenges facing governments who want to ensure citizens remain in control of the AI systems they deploy (cf. Alfrink et al., 2023).
  7. Why is this realist? Because the work looks at how some groups of people use particular forms of actually existing AI to do things to other people. The work also foregrounds the competing interests that are at stake. And it frames AI as neither fully autonomous nor fully passive, but as a thing that mediates peoples’ perceptions and actions.
  8. There are more examples besides this. But I will stop here. I just want to reiterate that I think we need a realist approach to the design of AI.

Participatory AI and ML engineering

In the first half of this year, I’ve presented several versions of a brief talk on participatory AI. I figured I would post an amalgam of these to the blog for future reference. (Previously, on the blog, I posted a brief lit review on the same topic; this talk builds on that.)

So, to start, the main point of this talk is that many participatory approaches to AI don’t engage deeply with the specifics of the technology. One such specific is the translation work engineers do to make a problem “learnable” by a machine (Kang, 2023). From this perspective, the main question to ask becomes, how does translation happen in our specific projects? Should citizens be involved in this translation work? If so, how to achieve this?

Before we dig into the state of participatory AI, let’s begin by clarifying why we might want to enable participation in the first place. A common motivation is a lack of democratic control over AI systems. (This is particularly concerning when AI systems are used for government policy execution. These are the systems I mostly look at in my own research.) And so the response is to bring the people into the development process, and to let them co-decide matters.

In these cases, participation can be understood as an enabler of democratic agency, i.e., a way for subjects to legitimate the use of AI systems (cf. Peter, 2020 in Rubel et al., 2021). Peter distinguishes two pathways: a normative one and a democratic one. Participation can be seen as an example of the democratic pathway to legitimation. A crucial detail Peter mentions here, which is often overlooked in participatory AI literature, is that normative constraints must limit the democratic pathway to avoid arbitrariness.

So, what is the state of participatory AI research and practice? I will look at each in turn next.

As mentioned, I previously posted on the state of participatory AI research, so I won’t repeat that in full here. (For the record, I reviewed Birhane et al. (2022), Bratteteig & Verne (2018), Delgado et al. (2023), Ehsan & Riedl (2020), Feffer et al. (2023), Gerdes (2022), Groves et al. (2023), Robertson et al. (2023), Sloane et al. (2020), and Zytko et al. (2022).) Elements that jump out include:

  • Superficial and unrepresentative involvement.
  • Piecemeal approaches that have minimal impact on decision-making.
  • Participants with a consultative role rather than that of active decision-makers.
  • A lack of bridge-builders between stakeholder perspectives.
  • Participation washing and exploitative community involvement.
  • Struggles with the dynamic nature of technology over time.
  • Discrepancies between the time scales for users to evaluate design ideas versus the pace at which systems are developed.
  • A demand for participation to enhance community knowledge and to actually empower them.

Taking a step back, if I were to evaluate the state of the scientific literature on participatory AI, it strikes me that many of these issues are not new to AI. They have been present in participatory design more broadly for some time already. Many of these issues are also not necessarily specific to AI. The ones I would call out include the issues related to AI system dynamism, time scales of participation versus development, and knowledge gaps between various actors in participatory processes (and, relatedly, the lack of bridge-builders).

So, what about practice? Let’s look at two reports that I feel are a good representation of the broader field: Framework for Meaningful Stakeholder Involvement by ECNL & SocietyInside, and Democratizing AI: Principles for Meaningful Public Participation by Data & Society.

Framework for Meaningful Stakeholder Involvement is aimed at businesses, organizations, and institutions that use AI. It focuses on human rights, ethical assessment, and compliance. It aims to be a tool for planning, delivering, and evaluating stakeholder engagement effectively, emphasizing three core elements: Shared Purpose, Trustworthy Process, and Visible Impact.

Democratizing AI frames public participation in AI development as a way to add legitimacy and accountability and to help prevent harmful impacts. It outlines risks associated with AI, including biased outcomes, opaque decision-making processes, and designers lacking real-world impact awareness. Causes for ineffective participation include unidirectional communication, socioeconomic barriers, superficial engagement, and ineffective third-party involvement. The report uses environmental law as a reference point and offers eight guidelines for meaningful public participation in AI.

Taking stock of these reports, we can say that the building blocks for the overall process are available to those seriously looking. The challenges facing participatory AI are, on the one hand, economic and political. On the other hand, they are related to the specifics of the technology at hand. For the remainder of this piece, let’s dig into the latter a bit more.

Let’s focus on translation work done by engineers during model development.

For this, I build on work by Kang (2023), which focuses on the qualitative analysis of how phenomena are translated into ML-compatible forms, paying specific attention to the ontological translations that occur in making a problem learnable. Translation in ML means transforming complex qualitative phenomena into quantifiable and computable forms. Multifaceted problems are converted into a “usable quantitative reference” or “ground truth.” This translation is not a mere representation of reality but a reformulation of a problem into mathematical terms, making it understandable and processable by ML algorithms. This transformation involves a significant amount of “ontological dissonance,” as it mediates and often simplifies the complexity of real-world phenomena into a taxonomy or set of classes for ML prediction. The process of translating is based on assumptions and standards that may alter the nature of the ML task and introduce new social and technical problems.

So what? I propose we can use the notion of translation as a frame for ML engineering. Understanding ML model engineering as translation is a potentially useful way to analyze what happens at each step of the process: What gets selected for translation, how the translation is performed, and what the resulting translation consists of.

So, if we seek to make participatory AI engage more with the technical particularities of ML, we could begin by identifying translations that have happened or might happen in our projects. We could then ask to what extent these acts of translation are value-laden. For those that are, we could think about how to communicate these translations to a lay audience. A particular challenge I expect we will be faced with is what the meaningful level of abstraction for citizen participation during AI development is. We should also ask what the appropriate ‘vehicle’ for citizen participation will be. And we should seek to move beyond small-scale, one-off, often unrepresentative forms of direct participation.

Bibliography

  • Birhane, A., Isaac, W., Prabhakaran, V., Diaz, M., Elish, M. C., Gabriel, I., & Mohamed, S. (2022). Power to the People? Opportunities and Challenges for Participatory AI. Equity and Access in Algorithms, Mechanisms, and Optimization, 1–8. https://doi.org/10/grnj99
  • Bratteteig, T., & Verne, G. (2018). Does AI make PD obsolete?: Exploring challenges from artificial intelligence to participatory design. Proceedings of the 15th Participatory Design Conference: Short Papers, Situated Actions, Workshops and Tutorial – Volume 2, 1–5. https://doi.org/10/ghsn84
  • Delgado, F., Yang, S., Madaio, M., & Yang, Q. (2023). The Participatory Turn in AI Design: Theoretical Foundations and the Current State of Practice. Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, 1–23. https://doi.org/10/gs8kvm
  • Ehsan, U., & Riedl, M. O. (2020). Human-Centered Explainable AI: Towards a Reflective Sociotechnical Approach. In C. Stephanidis, M. Kurosu, H. Degen, & L. Reinerman-Jones (Eds.), HCI International 2020—Late Breaking Papers: Multimodality and Intelligence (pp. 449–466). Springer International Publishing. https://doi.org/10/gskmgf
  • Feffer, M., Skirpan, M., Lipton, Z., & Heidari, H. (2023). From Preference Elicitation to Participatory ML: A Critical Survey & Guidelines for Future Research. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, 38–48. https://doi.org/10/gs8kvx
  • Gerdes, A. (2022). A participatory data-centric approach to AI Ethics by Design. Applied Artificial Intelligence, 36(1), 2009222. https://doi.org/10/gs8kt4
  • Groves, L., Peppin, A., Strait, A., & Brennan, J. (2023). Going public: The role of public participation approaches in commercial AI labs. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1162–1173. https://doi.org/10/gs8kvs
  • Kang, E. B. (2023). Ground truth tracings (GTT): On the epistemic limits of machine learning. Big Data & Society, 10(1), 1–12. https://doi.org/10/gtfgvx
  • Peter, F. (2020). The Grounds of Political Legitimacy. Journal of the American Philosophical Association, 6(3), 372–390. https://doi.org/10/grqfhn
  • Robertson, S., Nguyen, T., Hu, C., Albiston, C., Nikzad, A., & Salehi, N. (2023). Expressiveness, Cost, and Collectivism: How the Design of Preference Languages Shapes Participation in Algorithmic Decision-Making. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 1–16. https://doi.org/10/gr6q2t
  • Rubel, A., Castro, C., & Pham, A. K. (2021). Algorithms and autonomy: The ethics of automated decision systems. Cambridge University Press.
  • Sloane, M., Moss, E., Awomolo, O., & Forlano, L. (2020). Participation is not a Design Fix for Machine Learning. arXiv:2007.02423 [Cs]. http://arxiv.org/abs/2007.02423
  • Zytko, D., J. Wisniewski, P., Guha, S., P. S. Baumer, E., & Lee, M. K. (2022). Participatory Design of AI Systems: Opportunities and Challenges Across Diverse Users, Relationships, and Application Domains. Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems, 1–4. https://doi.org/10/gs8kv6

Democratizing AI Through Continuous Adaptability: The Role of DevOps

Below are the abstract and slides for my contribution to the TILTing Perspectives 2024 panel “The mutual shaping of democratic practices & AI,” moderated by Merel Noorman.

Slides

Abstract

Contestability

This presentation delves into democratizing artificial intelligence (AI) systems through contestability. Contestability refers to the ability of AI systems to remain open and responsive to disputes throughout their lifecycle. It approaches AI systems as arenas where groups compete for power over designs and outcomes.

Autonomy, democratic agency, legitimation

We identify contestability as a critical system quality for respecting people’s autonomy. This includes their democratic agency: their ability to legitimate policies. This includes policies enacted by AI systems.

For a decision to be legitimate, it must be democratically willed or rely on “normative authority.” The democratic pathway should be constrained by normative bounds to avoid arbitrariness. The appeal to authority should meet the “access constraint,” which ensures citizens can form beliefs about policies with a sufficient degree of agency (Peter, 2020 in Rubel et al., 2021).

Contestability is the quality that ensures mechanisms are in place for subjects to exercise their democratic agency. In the case of an appeal to normative authority, contestability mechanisms are how subjects and their representatives gain access to the information that will enable them to evaluate its justifiability. In this way, contestability satisfies the access constraint. In the case of democratic will, contestability-by-design practices are how system development is democratized. The autonomy account of legitimation adds the normative constraints that should bind this democratic pathway.

Himmelreich (2022) similarly argues that only a “thick” conception of democracy will address some of the current shortcomings of AI development. This is a pathway that not only allows for participation but also includes deliberation over justifications.

The agonistic arena

Elsewhere, we have proposed the Agonistic Arena as a metaphor for thinking about the democratization of AI systems (Alfrink et al., 2024). Contestable AI embodies the generative metaphor of the Arena. This metaphor characterizes public AI as a space where interlocutors embrace conflict as productive. Seen through the lens of the Arena, public AI problems stem from a need for opportunities for adversarial interaction between stakeholders.

This metaphorical framing suggests prescriptions to make more contentious and open to dispute the norms and procedures that shape:

  1. AI system design decisions on a global level, and
  2. human-AI system output decisions on a local level (i.e., individual decision outcomes), establishing new dialogical feedback loops between stakeholders that ensure continuous monitoring.

The Arena metaphor encourages a design ethos of revisability and reversibility so that AI systems embody the agonistic ideal of contingency.

Post-deployment malleability, feedback-ladenness

Unlike physical systems, AI technologies exhibit a unique malleability post-deployment.

For example, LLM chatbots optimize their performance based on a variety of feedback sources, including interactions with users, as well as feedback collected through crowd-sourced data work.

Because of this open-endedness, democratic control and oversight in the operations phase of the system’s lifecycle become a particular concern.

This is a concern because while AI systems are dynamic and feedback-laden (Gilbert et al., 2023), many of the existing oversight and control measures are static, one-off exercises that struggle to track systems as they evolve over time.

DevOps

The field of DevOps is pivotal in this context. DevOps focuses on system instrumentation for enhanced monitoring and control for continuous improvement. Typically, metrics for DevOps and their machine learning-specific MLOps offshoot emphasize technical performance and business objectives.

However, there is scope to expand these to include matters of public concern. The matters-of-concern perspective shifts the focus on issues such as fairness or discrimination, viewing them as challenges that cannot be resolved through universal methods with absolute certainty. Rather, it highlights how standards are locally negotiated within specific institutional contexts, emphasizing that such standards are never guaranteed (Lampland & Star, 2009, Geiger et al., 2023).

MLOps Metrics

In the context of machine learning systems, technical metrics focus on model accuracy. For example, a financial services company might use Area Under The Curve Receiver Operating Characteristics (AUC-ROC) to continuously monitor and maintain the performance of their fraud detection model in production.

Business metrics focus on cost-benefit analyses. For example, a bank might use a cost-benefit matrix to balance the potential revenue from approving a loan against the risk of default, ensuring that the overall profitability of their loan portfolio is optimized.

Drift

These metrics can be monitored over time to detect “drift” between a model and the world. Training sets are static. Reality is dynamic. It changes over time. Drift occurs when the nature of new input data diverges from the data a model was trained on. A change in performance metrics may be used to alert system operators, who can then investigate and decide on a course of action, e.g., retraining a model on updated data. This, in effect, creates a feedback loop between the system in use and its ongoing development.

An expansion of these practices in the interest of contestability would require:

  1. setting different metrics,
  2. exposing these metrics to additional audiences, and
  3. establishing feedback loops with the processes that govern models and the systems they are embedded in.

Example 1: Camera Cars

Let’s say a city government uses a camera-equipped vehicle and a computer vision model to detect potholes in public roads. In addition to accuracy and a favorable cost-benefit ratio, citizens, and road users in particular, may care about the time between a detected pothole and its fixing. Or, they may care about the distribution of potholes across the city. Furthermore, when road maintenance appears to be degrading, this should be taken up with department leadership, the responsible alderperson, and council members.

Example 2: EV Charching

Or, let’s say the same city government uses an algorithmic system to optimize public electric vehicle (EV) charging stations for green energy use by adapting charging speeds to expected sun and wind. EV drivers may want to know how much energy has been shifted to greener time windows and its trends. Without such visibility on a system’s actual goal achievement, citizens’ ability to legitimate its use suffers. As I have already mentioned, democratic agency, when enacted via the appeal to authority, depends on access to “normative facts” that underpin policies. And finally, professed system functionality must be demonstrated as well (Raji et al., 2022).

DevOps as sociotechnical leverage point for democratizing AI

These brief examples show that the DevOps approach is a potential sociotechnical leverage point. It offers pathways for democratizing AI system design, development, and operations.

DevOps can be adapted to further contestability. It creates new channels between human and machine actors. One of DevOps’s essential activities is monitoring (Smith, 2020), which presupposes fallibility, a necessary precondition for contestability. Finally, it requires and provides infrastructure for technical flexibility so that recovery from error is low-cost and continuous improvement becomes practically feasible.

The mutual shaping of democratic practices & AI

Zooming out further, let’s reflect on this panel’s overall theme, picking out three elements: legitimation, representation of marginalized groups, and dealing with conflict and contestation after implementation and during use.

Contestability is a lever for demanding justifications from operators, which is a necessary input for legitimation by subjects (Henin & Le Métayer, 2022). Contestability frames different actors’ stances as adversarial positions on a political field rather than “equally valid” perspectives (Scott, 2023). And finally, relations, monitoring, and revisability are all ways to give voice to and enable responsiveness to contestations (Genus & Stirling, 2018).

And again, all of these things can be furthered in the post-deployment phase by adapting the DevOps lens.

Bibliography

  • Alfrink, K., Keller, I., Kortuem, G., & Doorn, N. (2022). Contestable AI by Design: Towards a Framework. Minds and Machines, 33(4), 613–639. https://doi.org/10/gqnjcs
  • Alfrink, K., Keller, I., Yurrita Semperena, M., Bulygin, D., Kortuem, G., & Doorn, N. (2024). Envisioning Contestability Loops: Evaluating the Agonistic Arena as a Generative Metaphor for Public AI. She Ji: The Journal of Design, Economics, and Innovation, 10(1), 53–93. https://doi.org/10/gtzwft
  • Geiger, R. S., Tandon, U., Gakhokidze, A., Song, L., & Irani, L. (2023). Making Algorithms Public: Reimagining Auditing From Matters of Fact to Matters of Concern. International Journal of Communication, 18(0), Article 0.
  • Genus, A., & Stirling, A. (2018). Collingridge and the dilemma of control: Towards responsible and accountable innovation. Research Policy, 47(1), 61–69. https://doi.org/10/gcs7sn
  • Gilbert, T. K., Lambert, N., Dean, S., Zick, T., Snoswell, A., & Mehta, S. (2023). Reward Reports for Reinforcement Learning. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, 84–130. https://doi.org/10/gs9cnh
  • Henin, C., & Le Métayer, D. (2022). Beyond explainability: Justifiability and contestability of algorithmic decision systems. AI & SOCIETY, 37(4), 1397–1410. https://doi.org/10/gmg8pf
  • Himmelreich, J. (2022). Against “Democratizing AI.” AI & SOCIETY. https://doi.org/10/gr95d5
  • Lampland, M., & Star, S. L. (Eds.). (2008). Standards and Their Stories: How Quantifying, Classifying, and Formalizing Practices Shape Everyday Life (1st edition). Cornell University Press.
  • Peter, F. (2020). The Grounds of Political Legitimacy. Journal of the American Philosophical Association, 6(3), 372–390. https://doi.org/10/grqfhn
  • Raji, I. D., Kumar, I. E., Horowitz, A., & Selbst, A. (2022). The Fallacy of AI Functionality. 2022 ACM Conference on Fairness, Accountability, and Transparency, 959–972. https://doi.org/10/gqfvf5
  • Rubel, A., Castro, C., & Pham, A. K. (2021). Algorithms and autonomy: The ethics of automated decision systems. Cambridge University Press.
  • Scott, D. (2023). Diversifying the Deliberative Turn: Toward an Agonistic RRI. Science, Technology, & Human Values, 48(2), 295–318. https://doi.org/10/gpk2pr
  • Smith, J. D. (2020). Operations anti-patterns, DevOps solutions. Manning Publications.
  • Treveil, M. (2020). Introducing MLOps: How to scale machine learning in the enterprise (First edition). O’Reilly.

PhD update – June 2024

I am writing this final PhD update as a freshly minted doctor. On Thursday, May 23, 2024, I successfully defended my thesis, ‘Contestable Artificial Intelligence: Constructive Design Research for Public Artificial Intelligence Systems that are Open and Responsive to Dispute.’

I started the PhD on September 1, 2018 (read the very first update posted on that day here). So, that’s five years, eight months, 23 days from start to finish. It has been quite the journey, and I feel happy and relieved to have completed it. I am proud of the work embodied in the thesis. Most of all, I am thankful for the transformative learning experience, none of which would have been possible without the support of my supervisors Gerd, Ianus, and Neelke.

On the day itself, I was honored to have as my external committee members professors Dignum, Löwgren, van Zoonen, and van de Poel, professor Voûte as the chair, and Joost and Mireia as my paranymphs.

The thesis PDF can be downloaded at the TU Delft repository, and a video of the proceedings is available on YouTube.

Me, with a copy of the thesis, shortly before starting the layperson’s talk. Photo: Roy Borghouts.

Recent events

Reviewing my notes since the last update, below are some more notable things that happened in the past eight months.

  • I ran a short workshop on AI Pedagogy Through A Design Lens, together with Hosana Morales, at the TU Delft spring symposium on AI education. Read the post.
  • A story about my research was published on the TU Delft industrial design engineering website in the run-up to my defense on May 14, 2024. Read the story.
  • I updated and ran the fifth and final iteration of the AI & Society industrial design engineering master elective course from February 28 through April 10, 2024. A previous version is documented here, which I plan to update sometime in the near future.
  • I gave a talk titled Contestable AI: Designing for Human Autonomy at the Amsterdam UX meetup on February 21, 2024. Download the slides.
  • The outcomes of a design sprint on tools for third-party scrutiny, organized by the Responsible Sensing Lab, which took inspiration from my research, were published on December 7, 2023. Read the report.
  • I was interviewed by Mireia Yurrita Semperena for a DCODE podcast episode titled Beyond Values in Algorithmic Design, published November 6, 2023. Listen to the episode.
  • Together with Claudio Sarra and Marco Almada, I hosted an online seminar titled Building Contestable Systems on October 26, 2023. Read the thread.
  • I was a panelist at the Design & AI Symposium 2023 on October 18, 2023.
  • A paper I co-authored titled When ‘Doing Ethics’ Meets Public Procurement of Smart City Technology – an Amsterdam Case Study, was presented by first author Mike de Kreek at IASDR 2023 on October 9-13. Read the paper.

Looking ahead

I will continue at TU Delft as a postdoctoral researcher and will stay focused on design, AI, and politics, but I will try to evolve my research into something that builds on my thesis work but adds a new angle.

The Envisioning Contestability Loops article mentioned in previous updates is now in press with She Ji, which I am very pleased about. It should be published “soon.”

Upcoming appearances include a brief talk on participatory AI at a Cities Coalition for Digital Rights event and a presentation as part of a panel on The Mutual Shaping Of Democratic Practices And AI at TILTing Perspectives 2024.

That’s it for this final PhD update. I will probably continue these posts under a new title. We’ll see.

AI pedagogy through a design lens

At a TU Delft spring symposium on AI education, Hosana and I ran a short workshop titled “AI pedagogy through a design lens.” In it, we identified some of the challenges facing AI teaching, particularly outside of computer science, and explored how design pedagogy, particularly the practices of studios and making, may help to address them. The AI & Society master elective I’ve been developing and teaching over the past five years served as a case study. The session was punctuated by brief brainstorming using an adapted version of the SQUID gamestorming technique. Below are the slides we used.

Participatory AI literature review

I’ve been thinking alot about civic participation in machine learning systems development. In particular, involving non-experts in the potentially value-laden translation work from specifications that engineers do when they build their models. Below is a summary of a selection of literature I found on the topic, which may serve as a jumping-off point for future research.

Abstract

The literature on participatory artificial intelligence (AI) reveals a complex landscape marked by challenges and evolving methodologies. Feffer et al. (2023) critique the reduction of participation to computational mechanisms that only approximate narrow moral values. They also note that engagements with stakeholders are often superficial and unrepresentative. Groves et al. (2023) identify significant barriers in commercial AI labs, including high costs, fragmented approaches, exploitation concerns, lack of transparency, and contextual complexities. These barriers lead to a piecemeal approach to participation with minimal impact on decision-making in AI labs. Delgado et al. (2023) observe that participatory AI involves stakeholders mostly in a consultative role without integrating them as active decision-makers throughout the AI design lifecycle.

Gerdes (2022) proposes a data-centric approach to AI ethics and underscores the need for interdisciplinary bridge builders to reconcile different stakeholder perspectives. Robertson et al. (2023) explore participatory algorithm design, emphasizing the need for preference languages that balance expressiveness, cost, and collectivism—Sloane et al. (2020) caution against “participation washing” and the potential for exploitative community involvement. Bratteteig & Verne (2018) highlight AI’s challenges to traditional participatory design (PD) methods, including unpredictable technological changes and a lack of user-oriented evaluation. Birhane et al. (2022) call for a clearer understanding of meaningful participation, advocating for a shift towards vibrant, continuous engagement that enhances community knowledge and empowerment. The literature suggests a pressing need for more effective, inclusive, and empowering participatory approaches in AI development.

Bibliography

  1. Birhane, A., Isaac, W., Prabhakaran, V., Diaz, M., Elish, M. C., Gabriel, I., & Mohamed, S. (2022). Power to the People? Opportunities and Challenges for Participatory AI. Equity and Access in Algorithms, Mechanisms, and Optimization, 1–8. https://doi.org/10/grnj99
  2. Bratteteig, T., & Verne, G. (2018). Does AI make PD obsolete?: Exploring challenges from artificial intelligence to participatory design. Proceedings of the 15th Participatory Design Conference: Short Papers, Situated Actions, Workshops and Tutorial – Volume 2, 1–5. https://doi.org/10/ghsn84
  3. Delgado, F., Yang, S., Madaio, M., & Yang, Q. (2023). The Participatory Turn in AI Design: Theoretical Foundations and the Current State of Practice. Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, 1–23. https://doi.org/10/gs8kvm
  4. Ehsan, U., & Riedl, M. O. (2020). Human-Centered Explainable AI: Towards a Reflective Sociotechnical Approach. In C. Stephanidis, M. Kurosu, H. Degen, & L. Reinerman-Jones (Eds.), HCI International 2020—Late Breaking Papers: Multimodality and Intelligence (pp. 449–466). Springer International Publishing. https://doi.org/10/gskmgf
  5. Feffer, M., Skirpan, M., Lipton, Z., & Heidari, H. (2023). From Preference Elicitation to Participatory ML: A Critical Survey & Guidelines for Future Research. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, 38–48. https://doi.org/10/gs8kvx
  6. Gerdes, A. (2022). A participatory data-centric approach to AI Ethics by Design. Applied Artificial Intelligence, 36(1), 2009222. https://doi.org/10/gs8kt4
  7. Groves, L., Peppin, A., Strait, A., & Brennan, J. (2023). Going public: The role of public participation approaches in commercial AI labs. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1162–1173. https://doi.org/10/gs8kvs
  8. Robertson, S., Nguyen, T., Hu, C., Albiston, C., Nikzad, A., & Salehi, N. (2023). Expressiveness, Cost, and Collectivism: How the Design of Preference Languages Shapes Participation in Algorithmic Decision-Making. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 1–16. https://doi.org/10/gr6q2t
  9. Sloane, M., Moss, E., Awomolo, O., & Forlano, L. (2020). Participation is not a Design Fix for Machine Learning. arXiv:2007.02423 [Cs]. http://arxiv.org/abs/2007.02423
  10. Zytko, D., J. Wisniewski, P., Guha, S., P. S. Baumer, E., & Lee, M. K. (2022). Participatory Design of AI Systems: Opportunities and Challenges Across Diverse Users, Relationships, and Application Domains. Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems, 1–4. https://doi.org/10/gs8kv6

PhD update – September 2023

I’m back again with another Ph.D. update. Five years after I started in Delft, we are nearing the finish line on this whole thing. But before we look ahead, let’s review notable events since the previous update in March 2023.

Occurrences

  1. I presented our framework, Contestable AI by Design, at the annual NWO ICT Open conference, which, for the first time, had an entire track dedicated to HCI research in the Netherlands. It was an excellent opportunity to meet fellow researchers from other Dutch institutions. The slides are available as PDF at contestable.ai.
  2. I visited Hamburg to present our paper, Contestable Camera Cars, at CHI 2023. We also received a Best Paper award, which I am, of course, very pleased with. The conference was equal parts inspiring and overwhelming. The best part of it was meeting in-person researchers who shared my interests.
  3. Also, at CHI, I was interviewed about my research by Mike Green for his podcast Understanding Users. You can listen to it here. It is always good practice to try and lay out some of my arguments spontaneously live.
  4. In June, I joined a panel at a BOLD Cities “talk show” to discuss the design of smart city systems for contestability. It was quite an honor to be on the same panel as Eefje Cuppen, director of the Rathenau Institute. This event was great because we had technological, design, political, and policy perspectives. Several guests argued for the need to reinvigorate representative democracy and give a more prominent role to elected politicians in setting technology policy. A report is available here.
  5. In August, the BRIDE project had its closing event. This is the NWO research project that partially funded my Ph.D. The event was an excellent opportunity to reflect on our work together over the past years. I took the opportunity to revisit the work of Saskia Sassen on cityness and to think through some of the implications of my work on contestability for the field of smart urbanism. The slides are available at contestable.ai.
  6. Finally, last week, a short opinion piece that lays out the argument for contestable AI in what I hope is a reasonably accessible manner, was published on the TU Delft website.
Photo of Eefje Cuppen and I being interviewed by Inge Janse at the BOLD Cities talk show on June 22, 2023—photo by Tiffany Konings.

Eefje Cuppen and I being interviewed by Inge Janse at the BOLD Cities talk show on June 22, 2023—photo by Tiffany Konings.

Envisioning Contestability Loops

Throughout this, I have been diligently chipping away at my final publication, “Envisioning Contestability Loops: Evaluating the Agonistic Arena as a Generative Metaphor for Public AI.” I had a great time collaborating with Leon de Korte on an infographic of part of my design framework.

We took this infographic on a tour of Dutch interaction design agencies and conducted concept design workshops. I enjoyed returning to practice and sharing the work of the past couple of years with peers in practice. My friends at Eend wrote a nice blog post about it.

The analysis of the outcomes of these workshops forms the basis for the article, in which I explore the degree to which the guiding concept (generative metaphor) behind contestable AI, which I have dubbed the “Agonistic Arena” is a productive one for design practitioners. Spoilers: It is, but competing metaphors are also at play in the public AI design space.

The manuscript is close to completion. As usual, putting something like this together is a heavy but gratifying lift. I look forward to sharing the results and the underlying infographic with the broader world.

Are we there yet?

Looking ahead, I will be on a panel alongside the great Julian Bleecker and a host of others at the annual TU Delft Design & AI symposium in October.

But aside from that, I will keep my head down and focus on completing my thesis. The aim is to hand it in by the end of November. So, two more months on the clock. Will I make it? Let’s find out!