Worker Discretion Advised: Co-designing Risk Disclosure in Crowdsourced Responsible AI (RAI) Content Work

Paper Title

Worker Discretion Advised: Co-designing Risk Disclosure in Crowdsourced Responsible AI (RAI) Content Work

Publication Info

• Topic area: Risk disclosure mechanisms in crowdsourced Responsible AI (RAI) content work.
• Keywords: Responsible AI, crowdsourcing, risk disclosure, content moderation, worker well-being, task design, platform governance, psychological harm, co-design, sociotechnical systems.

Background and Problem

• Problem / challenge: Crowdsourced Responsible AI (RAI) content work often exposes workers to harmful material (e.g., graphic violence, hate speech) without adequate risk disclosure mechanisms. Current platforms provide limited support for worker well-being, and existing research focuses on employed moderators rather than crowdworkers.
• Significance: The lack of effective risk disclosure mechanisms can lead to psychological harm for crowdworkers, including burnout, anxiety, and PTSD, while also impacting data quality and participation in RAI tasks.
• Motivation and related work: Prior studies have documented the psychological risks of content moderation and structural power asymmetries in crowdwork, but have not sufficiently addressed how to design risk disclosure mechanisms that balance worker protection, task designer needs, and platform constraints. This paper addresses this gap by focusing on multi-stakeholder perspectives.

Solution

• Proposed approach: A co-design study involving 15 task designers, 11 crowdworkers, and 3 platform representatives to explore risk disclosure mechanisms in RAI content work.
• Novelty:
  1. Empirical insights into stakeholder perspectives on risk disclosure in crowdsourced RAI tasks.
  2. Analysis of sociotechnical tensions around risk disclosure, focusing on specificity, worker agency, and task designer agency.
  3. Design recommendations for accountable and transparent risk disclosure systems.
• Procedure and key techniques:
  • Conducted individual co-design sessions with stakeholders to explore preferences and tensions.
  • Focused on three design dimensions: warning specificity, worker agency, and task designer agency.
  • Used Figma-based prototypes and iterative feedback to surface design opportunities.
  • Applied reflexive thematic analysis to identify themes and tensions across stakeholder groups.

Results

• Concrete findings:
  • Workers prioritize clear, specific warnings and the ability to make informed decisions.
  • Task designers fear that detailed disclosures may deter participation but acknowledge ethical obligations.
  • Platforms aim to balance worker protection with scalability and liability concerns.
  • Workers reported psychological harm from inadequate warnings and emphasized the need for partial payment and feedback mechanisms.
• Advantage over baselines:
  • Proposes multi-stakeholder-informed designs, such as adaptive filters, AI-assisted warnings, and partial payment models, which address gaps in existing platform practices.
• Experiments / evaluation:
  • Co-design sessions with 29 participants (15 task designers, 11 workers, 3 platform representatives).
  • Explored design dimensions (specificity, agency) using prototypes and real-world examples.
  • Analyzed tensions and tradeoffs across task stages (e.g., sign-up, task completion, post-task feedback).
• Limitations and future work:
  • Limited access to platform representatives and potential selection bias in participant recruitment.
  • Future work should include perspectives from a broader range of platforms and roles, as well as longitudinal studies to evaluate proposed mechanisms.

Summary

This study investigates risk disclosure in crowdsourced Responsible AI (RAI) content work, where workers face significant psychological risks. Through co-design sessions with task designers, workers, and platform representatives, the authors identify tensions around warning specificity, worker agency, and task designer autonomy. Key contributions include empirical insights, an analysis of stakeholder tensions, and design recommendations such as AI-assisted warnings, partial payment models, and feedback mechanisms. These findings highlight the need for platforms to adopt more transparent and worker-centered risk disclosure practices while balancing ethical, operational, and scalability concerns.