Taxonomy of Collective Action in Algorithmic Systems
This post presents a taxonomy of ways in which users or collectives can resist algorithmic harm, organised by the collective goal they serve.
Auditing, Challenging, and Contesting Strategies: Making harms visible
These strategies aims to surface and contest harmful algorithmic behaviour through users everyday interactions with systems. The goal here for the users is to surface existing harm they come across in their everyday intractions of algorithms or better understand how an opaque model works through collectively theorising.
Folk theorisation [13] is the process by which users develop experiential understandings of how algorithmic systems behave. As an example, LGBTQ+ TikTok creators have built theories about how the For You Page interacts with identity [6], while queer creators in China use camouflaging strategies like hashtag substitution to navigate platform suppression [22]. Everyday algorithmic auditing [21] is the process by which users detect and interrogate harmful algorithmic behaviours through their everyday platform interactions. These observations, when shared publicly, can also generate collective evidence. The 2020 Twitter image-cropping case is an example where we can see the full cycle of auditing to action: users when interacting with the cropping algorithm noticed bias, shared evidence, and then built public pressure until Twitter abandoned the algorithm entirely [4]. [11] term this as algorithmic abandonment. Communities can also demand that platforms adopt less discriminatory algorithms [3], using the fact that multiple models often achieve similar accuracy but differ in their fairness metrics due to arbitrariness in training. Communities have also pursued legal contestation, challenging algorithmic discrimination in court [7, 25, 26].
Algorithmic Opt-Out Strategies: Preventing data misuse
The second category of strategies can be used by users of a system to prevent them from using user data for training or inference.
Users in data driven systems can leverage this data-dependency of AI systems through data leverage [27] through levers like data strikes (withholding data), data poisoning (deliberately manipulating data to degrade model performance), and conscious data contribution (redirecting data to better-aligned platforms). Data refusal [29] are strategies and practices which frame non-participation as a political act challenging the legitimacy of data collectors. An example for data strikes and data refusal can be seen through the 2023 Reddit blackout showing how these overlap in practice [17]. For users who may not be able to withdraw entirely from a platform, they can use different tools and systems to enact opt-out. Tools like Glaze [18] and Nightshade [19] apply perturbations to artwork that prevent generative AI models from learning artists’ styles. Another example is Fawkes [20] which protects photographs against facial recognition. Certain adversarial machine learning attacks can also be repurposed as a mechanism to enact opt-out such as unlearnable examples [10] which adds error-minimising noise to data so that models cannot extract meaningful patterns during training. Data defences [1] is a technique which injects adversarial prompts within text to prevent large language models from inferring identifying information. There are also various real world cases for platform migration/conscious data contribution such as queer women’s migration from TikTok to RedNote [14] and the mass departure of academics from Twitter/X to Bluesky [16].
Collective Intervention Strategies: Changing algorithmic behaviour
The third category includes strategies that actively alter how algorithmic systems behave, to get a favorable outcome.
Algorithmic Collective Action [8] provides a framework for how coordinated data modifications by groups of users can change model decisions. An example of coordinated collective action can be seen through the Decline Now campaign [23], where DoorDash drivers collectively declined low-payout orders to increase future payouts. Similarly, AdNauseam [9], a browser extension that automatically clicks all ads in the background, effectively poisons the data that targeted advertising algorithms rely on leading to change in behavior of targeted advertising towards users.
Decision Modification and Recommendation Strategies: Getting a better outcome
The fourth category focuses on obtaining better individual or collective outcomes from algorithmic systems through recourse.
Algorithmic recourse [12] provides users who receive unfavourable decisions with actionable recommendations for changing their outcome, typically through counterfactual explanations [28]. [5] extend this to online collective recourse, where coordinated modifications shift system behaviour for entire communities. An informal example is the widespread practice of optimising resumes to be “ATS-friendly” [15]. Model extraction techniques [2, 24] can also be repurposed to reconstruct approximate versions of deployed models, enabling communities to understand decision boundaries and develop more informed resistance strategies, though no community adoption has been documented to date.