Locating Risk: Task Designers and the Challenge of Risk Disclosure in RAI Content Work

Risk disclosure for crowdworkers is inconsistent and lacks standardized guidance: The paper found that existing transparency frameworks like model cards and datasheets focus on documenting AI models and datasets but fail to address how well-being risks should be communicated to the workers performing sensitive content moderation and red teaming tasks.

Task designers struggle to identify and communicate risks without clear workflows: Based on interviews with 23 task designers across academia and industry, the research reveals that in the absence of standard procedures, there is no consistent application of content warnings, consent processes, or other risk disclosure methods for workers exposed to violent, disturbing, or harmful content.

Current gaps in risk communication may harm both workers and AI development: The paper argues that inadequate risk disclosure not only threatens crowdworker well-being but also undermines the ethical integrity and technical quality of AI systems being developed, suggesting these issues are interconnected rather than separate concerns.

Responsible AI work creates significant but often undisclosed hazards: The study documents that crowdworkers regularly perform high-stakes tasks including labeling violent content, moderating disturbing text, and simulating harmful behavior—activities with documented risks to worker mental health that are frequently not properly communicated upfront.