Toward User-Driven Algorithm Auditing: Investigating Users' Strategies for Uncovering Harmful Algorithmic Behavior

Title of the Paper

Toward User-Driven Algorithm Auditing: Investigating users’ strategies for uncovering harmful algorithmic behavior

Paper Information

• Research Area: Human-Computer Interaction (HCI), Algorithm Auditing, and Social Computing
• Keywords: User-driven algorithm auditing, algorithmic bias, algorithmic harm, fair machine learning, fuzzy logic, collective work, qualitative methods, image search

Research Background and Problem

• Problem or Challenge:

  • Current algorithm auditing methods are predominantly expert-driven, which introduces cultural blind spots and limitations in addressing social dynamics, making it difficult to detect many biases and harms in real-world deployments.
  • Traditional auditing approaches overlook the involvement of actual users, potentially leaving bias issues undiscovered.
  • The potential for ordinary users to identify harmful algorithmic behavior through daily interactions remains underexplored.

• Significance:

  • Algorithmic bias can cause societal harm (e.g., stereotypes and misrepresentation), impacting fairness and social justice.
  • The ability of ordinary users to recognize harmful algorithmic behavior complements existing expert auditing methods and could help improve current algorithm auditing frameworks.

• Research Motivation and Related Work:

  • Recent studies suggest that ordinary users have the potential to identify and evaluate harmful algorithmic behavior, but how users can effectively perform algorithm auditing has not been thoroughly examined.
  • This research aims to explore the strategies of ordinary users to provide insights for designing more effective user-driven algorithm auditing tools in the future.
  • Against this backdrop, the study focuses on investigating how users identify and evaluate potentially harmful algorithmic behavior and the key factors influencing the user auditing process.

Proposed Solution

• Proposed Approach:

  1. Investigate how ordinary users evaluate algorithmic behavior through a three-phase study: including a status quo investigation task, a diary study, and collective workshops.
  2. Develop a three-phase model describing how users seek inspiration, construct meaning, and attempt remediation.
  3. Examine the main factors influencing user behavior: knowledge and beliefs, and platform affordances.

• Innovations:

  • Conducted the first in-depth study on the entire process of how users discover and understand algorithmic bias.
  • Developed a conceptual model describing the user process of bias discovery and meaning construction.
  • Proposed new directions for designing accessible and effective user-driven auditing tools.

• Implementation Steps and Techniques:

  1. Phase 1 (Think-Aloud Interviews): Designed individual tasks guiding participants to identify potential biases in image search results.
  2. Phase 2 (Diary Study): A 14-day study where participants documented instances of bias encountered during daily algorithm use, submitting 160 case reports.
  3. Phase 3 (Collective Workshops): Conducted four workshops to review submitted cases and explore collective understanding and analysis of biases.

Research Findings

• Specific Findings:

  • Users identified various types of algorithmic biases (e.g., age, gender, race, political, and religious biases).
  • Proposed a three-phase user auditing process model, including: inspiration seeking, meaning construction, and remediation actions.
  • Identified two main factors influencing user auditing behavior: knowledge and beliefs, and platform affordances.

• Comparison with Existing Solutions:

  • User-driven auditing is more grounded in real-world scenarios and capable of identifying biases within social dynamics and cultural contexts compared to expert-driven auditing.
  • Collective discussions help address the limitations of individual user perspectives, leading to more comprehensive bias understanding.

• Experimental or Evaluation Results:

  • Participants submitted 160 bias-related case reports, most of which were linked to algorithmic mechanisms and deemed potentially harmful.
  • The workshop process demonstrated that collective discussions effectively validated biases and generated more detailed remediation plans.

• Limitations and Future Directions:

  • Limitations include the study’s focus on image search, excluding other types of algorithms.
  • User-driven auditing methods may be influenced by cognitive biases, necessitating strategies to mitigate these biases.
  • Proposed the design of user-driven auditing platforms to balance team diversity and address information gaps.
  • Suggested exploring ways to enhance public awareness and transparency when algorithms cannot be fully corrected.