"I Don't Think RAI Applies to My Model'' -- Engaging Non-champions with Sticky Stories for Responsible AI Work

Most responsible AI tools fail to reach the practitioners who need them most. The research found that data scientists often dismiss RAI checklists, templates, and governance processes as irrelevant bureaucratic tasks rather than essential to their work, meaning current approaches primarily engage only those already motivated to care about ethics ("RAI champions").

"Sticky stories" — concrete narratives of real ML harms — significantly outperform generic information in changing practitioner behavior. When researchers tested stories designed to be specific, severe, surprising, and diverse against regular stories, practitioners spent more time identifying potential harms and recognized a broader range of risks, suggesting that emotionally resonant examples are more effective than abstract guidelines.

Practitioners are desensitized to widely circulated media about AI failures. The paper notes that data scientists have become numb to typical news coverage of AI harms, which is why the research specifically designed "sticky stories" to be unexpected and novel rather than relying on familiar cautionary tales.

Responsible AI engagement requires meeting practitioners where they are. Rather than imposing top-down processes, the research suggests that effective RAI work must address practitioners' perception that ethical considerations don't apply to their specific models, requiring tailored narratives that demonstrate direct relevance to their work.

Scale matters for responsible AI adoption. The researchers used AI systems to generate and evaluate sticky stories at scale, confirming this approach could potentially be deployed broadly across organizations rather than relying on one-off interventions.