TLDR: This paper introduces the first theoretical framework for Algorithmic Collective Action (ACA) involving multiple groups. It explores how different collectives, varying in size and goals, can coordinate to bias AI classifiers. The research provides quantitative insights into how factors like group size, signal uniqueness, and alignment of goals influence both individual and overall success, offering a crucial step towards understanding complex, real-world collective efforts to steer AI.
As artificial intelligence systems become more integrated into our daily lives, influencing everything from recommendations to critical decisions, the idea of users actively shaping these systems has gained traction. This concept, known as Algorithmic Collective Action (ACA), involves groups of users coordinating changes to shared data to steer AI models in a desired direction. Traditionally, much of the research in ACA has focused on scenarios where a single, unified group acts. However, real-world collective efforts are rarely monolithic; they are often decentralized and fragmented, involving multiple groups with varying sizes, strategies, and specific goals, even if they share broader objectives like climate justice or gender equality.
A New Framework for Multiple Collectives
A groundbreaking new research paper, titled Algorithmic Collective Action with Multiple Collectives, introduces the first theoretical framework to address ACA in settings with multiple collectives acting on the same AI system. The authors, Claudio Battiloro, Pietro Greiner, Bret Nestor, Oumaima Amezgar, and Francesca Dominici, delve into how these diverse groups can work together to influence classification models. Their focus is on how multiple collectives can “plant signals” – essentially biasing a classifier to learn a specific association between an altered version of data features and a chosen set of target categories.
The paper explores two main ways collectives can implement their strategies:
- Feature-label strategies: Here, collectives have the ability to modify both the input data (features) and the desired outcome (labels).
- Feature-only strategies: In this scenario, collectives can only alter the features of the data, while the labels remain fixed or are determined by other means.
Measuring Success and Key Influences
To understand the impact of these collective actions, the researchers define measures for both individual and global success. Per-collective success gauges how well each individual group achieves its specific objective. For global success, they propose two distinct metrics:
- “No One is Left Behind” (Smin): This egalitarian metric focuses on the success of the least successful collective, aiming to ensure that all groups benefit.
- “The Bigger the Better” (Sw): This metric averages the per-collective successes, weighting them by the size (or “mass”) of each collective, reflecting a proportional impact.
The framework also identifies several key parameters that influence the success of collective action:
- Collectives’ masses: The size or proportion of users belonging to each collective.
- Per-collective uniqueness: How rare or distinct the signal a specific collective is trying to plant is under the baseline conditions.
- Global uniqueness: How much the signals planted by different collectives overlap with each other.
- Target alignment: The total mass of other collectives that share the same target goal as a given collective.
- Suboptimality gap: A measure of how much the target label differs from what the baseline data distribution would naturally suggest.
The quantitative results reveal interesting trade-offs driven by the interplay of these factors, highlighting how group sizes and the alignment of their goals are crucial for determining the overall success of an action. The paper discusses how these factors contribute to a collective’s ability to influence the AI, considering aspects like the rarity and difficulty of planting a signal, as well as the competitive landscape among different collectives.
Real-World Application: Climate Adaptation in Cities
To illustrate the practical relevance of their framework, the authors present a compelling use case in climate adaptation. Imagine a city that uses an AI text classifier to identify necessary interventions based on neighborhood records, including community forum discussions. Grassroots neighborhood associations, acting as collectives, could coordinate to change their shared data to steer the classifier towards interventions they deem most necessary, such as requesting rain gardens or bioswales based on documented flooding. This could involve feature-label strategies (e.g., explicitly requesting an intervention) or feature-only strategies (e.g., only documenting issues if the city’s system doesn’t allow direct label edits).
In such a scenario, city-wide success could be measured by either the “No One is Left Behind” metric, emphasizing equitable outcomes for all neighborhoods, or the “The Bigger the Better” metric, reflecting the average success weighted by the size of each neighborhood association.
Also Read:
- Beyond Bias: Integrating Intersectional Feminist Theory into Algorithmic Fairness
- Unpacking AI’s Influence: How Language Models Persuade and Amplify Bias
Future Directions
This theoretical framework is a foundational step. The authors suggest several future research directions, including a deeper characterization of the “critical mass” needed for success in multi-collective settings, generalizing existing statistical frameworks, and exploring scenarios where collectives aim to erase or unplant signals. They also highlight the importance of studying mixed-objective settings and situations where collectives have heterogeneous capabilities, meaning some might use feature-label strategies while others are restricted to feature-only approaches.
In conclusion, this work provides a rigorous theoretical foundation for understanding Algorithmic Collective Action when multiple groups are involved. By quantifying the roles of collective sizes, signal uniqueness, and goal alignment, it paves the way for a more holistic and realistic treatment of how users can collectively steer the AI systems that shape their world.
