ALGORITHMIC COLLECTIVE ACTION

FAQ

• How can I cite this work?

  If you find this website useful, and decide to use its contents or the data it contains, you can do it for non-commercial purposes. We only require that you cite us and link back to the website.

• Who contributed to the tracker?

  Our heartfelt gratitude goes to our contributors for suggesting cases, ideas or pointing out typos in the website: Simon Zilinskas, Dorothee Sigg, Celestine Mendler-Dünner, Dariia Haryfullina, Simon Chignard. This repository is richer thanks to your inputs!

  You can be the next one to contribute to this collective effort by reaching out here.

• Who is this website for?

  This website is targeted at both civil society and the machine learning community, or anyone looking for inspiration or examples to inform actions and theories.

  For this reason, we have put some effort in making the website as accessible as possible. If you find areas for improvement, if you'd like to get involved, or if something is unclear, do not hesitate to reach out.

• How can I contribute?

  We envision this website as a participatory initiative. You are welcome to contribute cases or ideas through our contact form.

  If you found what seems to be an error, feel free to let us know here as well.

• What is this website about?

  We propose a taxonomy and a tracker useful both to the machine learning community and grassroots collectives looking for inspiration or examples to inform theories and actions. This work can be seen as a (non-exhaustive, and always-evolving) library of real, inspiring and documented examples that have occurred since the beginning of the 2000s. The objective is to give traction to the field of Algorithmic Collective Action, and to show that it has real practical applications that further motivate theoretical work.

• What is this website not about / What are its shortcomings?

  We wish to underline three key choices and limitations of this work.

  First, we focus on Algorithmic Collective Action and the added value of collectives in the digital world. This is a deliberate focus, and we believe it is complementary to other efforts such as the AI Resist List, or the Worker mobilizations in Culture tracker by focusing on a specific and complementary avenue for action.

  Secondly, this work is not exhaustive: our intention is to draw a first map of the field. Each case is presented rigorously with relevant sources, but no case has been the subject of in-depth case studies, nor has it directly involved the organizers. We intend to update the website with community contributions.

  Finally, the cases were collected primarily by people living in Europe. We thus acknowledge the associated biases: it is possible that cases from Europe and North America are over-represented compared to cases in South America, Africa, the Middle East, China and Oceania. We hope this work can reach diverse communities and encourage them to help us make the list more representative.

• Who built this website?

  This website was built and is maintained by Léo le Douarec, as part of his research internship at the ELLIS Institute Tübingen under the supervision of Celestine Mendler-Dünner. Dariia Haryfullina helped draft content and draft the website. She also created the logo.

• Is this website AI-generated?

  TL;DR: Only the website code, not the content.

  Detailed disclosure on AI usage: this website was co-created using an OpenCode coding agent for the HTML/CSS/Javascript web development and to help with translation. The collection of real-world cases, their annotation, the writing of their summaries, and all surrounding work was not AI-generated.

• What is in the taxonomy?

  Algorithmic Collective Action Knowledge & Coordination Real cases & related research

  The taxonomy (presented in the graph on the landing page) describes the field of Algorithmic Collective Action along two main axes: the AI Lifecycle, and the level of knowledge one possesses about the targeted algorithm.

  Data Collection: At this step, collectives can either choose to give their data, not give it at all, or to sell it. These methods require little to no knowledge of the system.

  Training: actions can target the training of an algorithm (or even pre/post-training for LLMs). There, the taxonomy differentiates actions based on the collective's intent. Either the aim is to harm the system (Data Poisoning), or the intention is to make a system better, and the action is then considered as Data Mending. Either way, these strategies require the most information possible from the algorithm. A significant portion of the theoretical academic literature on Algorithmic Collective Action focuses on the (re)-training aspect.

  Deployment: here, we consider a fixed model that has been deployed and with which the collective will interact, for example sending proxies or probes to understand its behaviour and limits. These strategies require less information than actions at the training stage.

  Evaluation: model providers often request user feedback to assess the performance and adequacy of their models. This evaluation phase is an avenue for collectives that would like to steer the model's development in a given direction by manipulating the feedback in a strategic way. All the collective needs is a well-defined objective to push for.

  Algorithmic Collective Action does not exist in the void by itself. It requires knowledge about the algorithm(s), coordination on the ground, and motivations for action. The tracker thus presents additional filters to classify cases depending on their knowledge acquisition techniques, coordination methods and themes of action. A dedicated inductive taxonomy was produced from the cases for these filters.

  This allows you to search cases based on how the collectives coordinate or share knowledge, what is the theme / reason for their action, what groups of people or acting, where, and when.

  The main focus of the tracker are the 126 real-life cases of ACA. There are also 47 research papers or initiatives that theorize the effectiveness of ACA in certain settings without confronting the model to reality, or study gig workers and their working conditions. We think all these contributions are valuable, and they are provided as separate tabs in the tracker.

• How was the taxonomy built?

  This taxonomy was mainly built using an inductive approach: we started from the list of cases, and tried grouping them by similarities to build a taxonomy that best describes the observed landscape. We also based ourselves on a few related taxonomies, especially of collective action in general for Traditional Collective Action, and of resistance against algorithms as well as Data Leverage by Nicholas Vincent for ACA. By iterating, annotating examples, and growing the list of cases, we converged towards the current taxonomy.

  The examples were collected from multiple already-existing lists, as well as personal examples. Already-existing lists used as a starting point include Dorothee Sigg's GitHub repository, talks from the ACA Workshop at NeurIPS, DAIR's Luddite Lab Resource Hub, Data & Society's newsletter, and a list built by Simon Chignard shared during the course "Data and Algorithms for Public Policy" at Sciences Po Paris. In the initial version of the tracker, cases were collected by Léo le Douarec. He was supported in this collection by Celestine Mendler-Dünner, Dariia Haryfullina and Simon Chignard.

  The design of the tracker was inspired by the ODAP, and the tracker on Worker Mobilizations around AI in Arts, Culture, and Media.

• What is Traditional Collective Action?

  We use the concept of "Traditional Collective Action" to put Algorithmic Collective Action (ACA) in context: ACA is one of the many ways collectives can act in algorithmically-mediated systems. This work aims to complement already-existing work on collective action against algorithms, and more generally, collective action.

  Throughout history, collectives have acted in various ways and scholars (sociologists, anthropologists)... have worked to understand these dynamics. In the realm of algorithmic systems, we distinguish two main avenues that people have traditionally used against algorithms:

  1. Institutional Action

  This relates to actions that leverage an institution and its processes to ensure a collective win.

  • It can be the judiciary institutions: With the rise of generative AI, legal cases have sprouted, especially on Intellectual Property and Copyright of training data. Other legal cases exist with non-generative AI, such as the CNAF Risk Scoring algorithm.
  • It can also be Collective Bargaining by trade unions negotiating with management, to reach a collective agreement on how organizations should use algorithms in the workplace, or how to respect workers’ requests (as in the Writers Guild of America’s strike over a labor dispute with the Alliance of Motion Picture and Television Producers).
  2. Grassroots Movements

  This depicts actions led by collectives who organize rather autonomously and exert pressure via various ways: protests, strikes, open letters, boycotts or conscious consumption campaigns, to name a few examples.

  All these actions constitute what we call “Traditional Collective Action”. The point of this distinction with Algorithmic Collective Action is that these levers are not specific to algorithmic systems. What's new with Algorithmic Collective Action is the statistical and dynamic learning nature of some of today’s algorithms. This reveals new levers (namely user-generated data), that require a collective to be as efficient as possible.

  More Examples

  Interactive taxonomy of Traditional Collective Actions in algorithmically-mediated systems.